>There's an important distinction here. Anthropic isn't actually spending more cash on AI than salaries. It's valuing its internal AI consumption at retail prices.
If virtually any other company in the world said "we have internal AI servers and we let our employees use them" ... the company would measure that as they would any other internal expenditure: by how much it costs them. You blow $50k a year on servers running your LLMs? It cost you $50k.
Only a company like Anthropic, desperate to make news and help their IPO, would measure by the cost they charge customers. It's as ridiculous as Microsoft counting every employee's use of MS Word ... as if it had paid the full consumer price of each copy.
> Also, Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
They have infinite money to burn so they aren't going to optimize.
I've seen that trivial changes to workflows can token usage by 30%. Repeat this a few times and my token usage has plummeted over the last few months.
Only outputting test failures instead of having agents read test run output in real time. Not having build logs get fed into context. Projects like UCL (unified context layer) so agents don't rediscover the same facts about a code base again and again.
Giving tools agent output modes so they only print out the minimal info needed for an agent to get work done.
When I hear a company has an absurd AI spend I assume they aren't even trying to reign it in.
> They have infinite money to burn so they aren't going to optimize.
They also kind of ("kind of" being the important clarifier here) can't optimize. See this is the trend with agentifying companies and the achilles heel for the frontier LLMs. We're starting to see diminishing returns on their capabilities, which means the rational enterprise is going to simply use the cheapest most efficient model for the task at hand. In turn we can just use an agent broker to decide which model is best and cost the least. Anthropic (today) has no incentive to do that other than to just use their own, best model for its own work.
> They have infinite money to burn so they aren't going to optimize.
IPOs aren't the magical moment that they were in decades past. You get to answer to a bunch of retirement fund managers and "the market", meaning less-savvy retail investors. You have to open up your books.
It seems like Anthropic, OpenAI, etc. are doing an IPO because they are literally out of other options. Older IPOs like Facebook felt different, like the hard part had been done, and they just had to scale. There's a difference in selling something because you want to, versus selling something because you have to. Facebook was just relying on the idea that they were a platform that more people would inevitably sign up for. For the most part, they did. It seemed obvious. They likely could have continued on by simply operating privately, but the IPO gave them capital to dominate the market.
Anthropic and other hyperscalers cannot continue on privately. They're facing a systemic shortage of capital to achieve their goals. They need the infrastructure, and there's not enough capacity to make enough for all of these companies to survive.
> Older IPOs like Facebook felt different, like the hard part had been done, and they just had to scale.
Facebook were actually forced to IPO, because they had too many shareholders to remain private. They were super generous with allowing early employees to sell equity, and the laws required them to list once they passed 500 (i think) external shareholders.
This got changed in later legislation, which is part of the reason that companies like Stripe have been able to remain private for so much longer.
That being said, I agree with basically everything else in your comment, including the feelings around this.
LLMs (and Claude more specifically) are wonderful products, but I'm not convinced that they can justify their valuations over time.
Anthropic has ~$146.6 billion in cash on hand (Sept 2026). It's not out of options at this point. That's the second largest pile of cash on hand of any corporation in the US outside of Berkshire Hathaway and Google.
The cheap money days are likely over or rapidly diminishing. The next plot of cheap money to add to their pile, is public investors (if they hurry). They can raise $75-$100 billion that way, pushing their pile well over $200 billion. One assumes based on their obligations and expenditures projected over the coming 3-5 years, they'll need it.
I've read that the build out of this current wave of AI infra is going to be about $700 billion [0]
I don't have the math behind that number, but it sounds somewhat in-line with what we know the market to be. You have a few companies seeking trillion+-dollar valuations at IPO, and the spends from established players make that number sound plausible.
Given that the infra costs are going to remain stubbornly high for the coming 3-5 year period you specified, I don't know that even $200 billion is really that much in the grand scheme of things. If things keep going at more-or-less the same pace they have been with advances, and SV pulls their heads out of their asses about getting data centers built out, maybe.
Human in the loop is 1/5th the token usage vs fully autonomous agents.
Agents swarming will 5x your usage, or worse.
The communication overhead of summarizing and status reporting between workers will eat up tokens faster than doing the work will.
Large monolith code bases with multiple layers of abstraction (Typescript to JSON schema, prisma, etc) will eat up more context than the actual work, by far. IMHO this same rule applies for human engineers as well, one of the reasons why I generally oppose too many layers of abstractions/middleware.
On large complex highly abstracted code bases, up to 90% of token usage will be input tokens.
Generally cached, generally cheaper, but not free.
The more you can constrain a problem (talking to black boxes with well defined contracts) the less the agent has to work.
So, you know, good software engineering principles in general.
Thanks! if you can think of any other large wins that you didn't mention, that'd be great; but those 5x numbers... that's brutal. I can understand the communications overhead of agent swarms; I'd love to know where the 5x is mostly going between human in the loop and not.
Towers of madness... that's absolutely perfect. I'm stealing that. This has happened so many times that it lives as a concept in my head but it never got a label. Thank you.
Like, it's wild to me seeing majority Chinese ML teams for years in the Bay and elsewhere that suddenly in the LLM space they are only capable of competing by stealing the hard stolen goods of our frontier US developers.
And people lap it up. And every Chinese advance - also made public, usually through papers - is implemented immediately by Western companies. Some of these are big steps forward and could have possibly been patented but they just share.
And the naming... 'Anthropic' & 'OpenAI' are such misnomers, that's not an accident.
Largely bad pathways that aren't being caught in time.
Inefficient commands, for example I launched a bunch of agents for a task and for every command they'd first try it inside a sandbox, wait for the command to fail (sometimes that meant a timeout!) then retry outside the sandbox.
Now do that hundreds of times.
Or just crap like type checking the entire repo again and again.
If you have really tight rein on your agents you can prevent these sorts of issues, but you won't see these issues unless you are carefully monitoring what they are doing to the level of watching cli output scroll by.
Of course at some point the agent.md file needed to keep the stupid away gets so long that adherence drops.
You also end up with just stupid shit. I gave Sol too much free reign and it decided to update my DB schema to save stack traces to a DB.
Code which it then wrote a bunch of tests for. And then used a browser use agent to manually test.
It’s not opportunity cost. It would be opportunity cost if their employees not using it would lead to additional revenue. However, if Inwas to go to Anthropic and double my token usage Anthropic wouldn’t decline my money. If Anthropic’s employees would stop using Claude, their revenues wouldn’t increase by an equivalent amount.
Their revenues are not currently supply constrained.
It’s just a cost cost. If their employees stopped using these services they would simply need less infrastructure, and therefore save on the cost to build and run that infrastructure. It wouldn’t translate into an increase in revenue.
> It would be opportunity cost if their employees not using it would lead to additional revenue. However, if Inwas to go to Anthropic and double my token usage Anthropic wouldn’t decline my money. If Anthropic’s employees would stop using Claude, their revenues wouldn’t increase by an equivalent amount.
I don’t agree. This is determined by elasticity of demand. If they have more compute to offer, that means they could offer higher subscription limits, more features like Duo to more people, better API pricing, and better enterprise contracts. This leads to some degree of increased consumption and more revenue. The degree is determined by how responsive customers are to pricing and features and performance. It’s not zero, so there is some amount that revenue would increase. We can debate the coefficient and the amount of compute we’re talking about here, but the premise is correct.
I don't buy it. A company like Anthropic is not going to just walk away from revenue. If external demand were really that high, then they should take the money first. And not just for the revenue; customer demand being that high is also an indication that the product is already great as-is so they stand to improve their bottom line by easing off on the R&D spend.
you could not be more wrong, you cannot IPO if you say product is good enough. The whole point, the bet why its worth trillions is because they will reach AGI and "singularity" first, so the gap vs open source will only increase. If you say product is great as is, then you will be caught up and game over
They are making a deliberate trade-off decision
token use for their own
vs
token use for the rest of the world.
while Google is making the opposite bet
In order to correctly make that decision, they must use the opportunity cost of selling that token. I'm pushing back at OPs claim that Anthropic must use internal cost to calculate that trade-off. They absolutely should use external selling price.
Well, that's the other side of it. If we say they are making a trade-off decision, that means that they are choosing not to sell to paying customers. Who are these paying customers that Anthropic is turning down? I haven't heard any headlines saying they've instituted token quotas or started putting potential new customers on a waiting list.
It's complete bologna. Basic GAAP accounting eliminates intra-company service consumption/sales. Looking at cloud compute, say, Amazon "bills" departments for AWS cloud compute consumption, but reporting internal cloud compute consumption as a thing they were doing assuming the rates charged to their least efficient customer would be beyond meaningless.
I don't understand why anyone is even talking about the cost the tokens when the anthropic is on both sides.
They pay for their own tokens and then they earn it back, net profit is negative, it's an expense but the expense is much smaller than the booked revenue.
They could be pricing each token at a billion dollars or a trillion tokens for a cent and the GAAP accounting wouldn't change.
They are aware of the opportunity cost of their tokens and the their staff are 'laying golden eggs' right now, if there is anyone making use of tokens, it's them.
Even tokens on the floor are probably well spent, given the economic leverage they have.
Not sure that's too relevant when the difference between revenue and valuation differs by 3 orders of magnitude (ie company is priced explicitly for growth).
It's supply constrained in terms of global hardware production capacity, but that's not how it works for any given company. If Anthropic wanted more GPUs for customers paying them enough money to outbid the next guy in the market for the finite number of GPUs then they would do that and get a higher percentage of the finite number of GPUs.
So the cost of using more internally is still their own cost, i.e. the price of being the high bidder on more compute resources, not the retail price they charge to customers.
You're missing the part where they can buy GPUs that would otherwise have gone to someone who is neither them nor their customers, for less than the amount that they charge their customers to use them.
It’s supply constrained for now, the moment companies like Anthropic have to switch to profit-making-mode, demand will dip as the prices skyrocket imo.
> Also, Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
That's API-list prices. They have an 80% margin over COGS.
Cost per equivalent output is dropping 10x every 18 months -> so in 3-ish years, if the trend continues, that will be $3k per employee.
Not bad. A lot of companies are already spending about that much or more on IT per employee.
Especially since that $3k bakes in an 80% margin. It's about $600 per year if you DIY at scale - which is about $50 per month - but at the $3k price, hardly anyone is going to.
In the West old-age dependency ratio (retirees per working-age adult) is climbing roughly 0.5 to 0.7 percentage points per year.
We've got to get 0.7% more productive for employees to keep the same quality of life (unless retirees get screwed, and since they vote and make up a huge portion of the voting block, I wouldn't count on that happening).
Further, more and more of the population has to transition into into elder care.
Unless we hit 3%+ GDP growth, us workers are probably going to see declines in quality of life (and we're probably not going to get those levels of growth)...
> We've got to get 0.7% more productive for employees to keep the same quality of life
There's countless ways compensate for this gap, some of which are regulatory, others merely obvious (like the lack of a multi-trillion AI bubble in the future), the assumption that technological progress continues as modern levels (ie "same quality of life") with the changes that may entail. Ofc, the demographic shift (as you mentioned) changes what "quality of life" even means, making it even less deterministic.
> Unless we hit 3%+ GDP growth, us workers are probably going to see declines in quality of life (and we're probably not going to get those levels of growth)
> A clear example is Google. They can easily give all compute to Deepmind, yet majority is allocated to be sold at market rate.
Google is in the business of buying compute and renting it out. To the extent that their internal use reduces revenues from that business, there is an opportunity cost.
It depends on your assumptions. When Google is in "space crunch" mode, there are opportunity costs. Otherwise, there aren't, because you could do both.
You could also look at the opportunity cost of not letting Deepmind do quite as much research. How much is additional research worth on the margin? There will be diminishing returns.
They also offer $300,000 in compute spend (allegedly) for $200 a month to anyone who isn't employed by a large corporation so their pricing should be taken with a grain of salt.
> If you use 100% of the quota every week it's closer to $2000 at API rates.
I don't know where you got that from but the only time I maxed my $20 plan for a week I made the calculation that I had spend close to $200 in a single week. So maxing a $200 plan for a month should give you $8000 worth of token at API price. Obviously it's not $300k but it's still very far from the retail price.
> but the only time I maxed my $20 plan for a week I made the calculation that I had spend close to $200 in a single week. So maxing a $200 plan for a month should give you $8000 worth of token at API price.
$200 of usage on a $20/mo plan is a 10X multiplier.
The $200/month plan has 10X the usage of the $20/month plan.
So $200 * 10X = $2000/month. I don't know how you got to $8000.
Recent analysis by industry analysts like Semi Accurate have measured Sol 6.1 on the $200/month plan to be around $2000 of API equivalent use.
I get a LOT of mileage out of my $20 plan. It's at least $500+ a week at API rates. I sometimes wonder if there's something wrong with their billing system.
On the $200* consumer plan, according to "npx ccusage", I'm using $11,000 per month at API rates. Possibly significantly more, because I was previously using the $100 plan so that only reflects about 1 week of the $200 plan.
It's not just the LLM costs, to keep your agents busy you need to move them of your workstation onto a server/cloud. I currently have about 16 claude code agents working on different bugs, uses so much resources that my macbook pro m5 max becomes slow to respond. Using cloud session would free up my macbook for other work.
Anthropic recently handed out free cloud resources (it's still free, guess they want you to get used to the fact that cloud is not free). I went through $250 cloud budget in less than 2 days.
I'm impressed you have a workflow that can keep 16 separate agents fed without the whole thing collapsing into nonsense. Even running full agent teams I can't think of ways to do that without being intentionally inefficient
I let one agent categorize all bugs, then have it decide which clusters of bugs should be addressed together, and which ones can be addressed in parallel.
Technically it is... According the original vision for workstation, a computer with >1MB RAM, >1 megapixel of screen resolution, >1 MIPS and costing under 1 Mega cents (10K dollars).
My laptop is shockingly competitive with my quite nice workstation.
(IMO the really impressive part is the thermals. My laptop can dissipate some 45W on a sustained basis, and my workstation takes quite a bit more than 45W to outperform my laptop.)
6 years ago i had a lenovo thinkpad p1, branded by lenovo as a portable workstation. Today i have macbook pro m5 max, much faster and more powerful than the lenovo, so why not call it a workstation?
So every business should measure any client fungible resource that way? Should a bank put a revenue number to any capital held in excess of regulatory requirements, for example?
Excess capacity sitting idle is the cardinal sin in modern economy.
I was merely pushing back at the insinuation that internal costs should be used to measure internal resource use.
No. The costs should always be external market price because that's where you are making the correct trade-offs.
Bank reserves, Compute etc are trade-off decisions.
Buffett vs Mayoshi
Anthropic vs Google
are being played out in real-time for us. Pretty soon you'll know which trade-off was better.
Zero excess capacity makes things quite unstable and estimating what revenue could potentially be had isn't necessarily easy or so reliable (and what about future revenues from using capacity internally vs revenued now?).
There is a huge difference between Anthropic and Google dog-fooding: Anthropic does not own or build data centers, they rent (as far as all the reporting on this I've read). Anthropic employees dog-fooding with Claude tools isn't "internal", its most likely being run on external AI data center compute, which they have to pay for, there's just no margin on it. It might even be run out of Google's data centers.
Is there any reason to believe that if this computer wasn’t used by an anthropic employee that it would be sold at retail price? That would imply that Anthropic’s computer is 100% utilized at all times, with no slack - something highly unlikely.
If it were opportunity cost, that would imply that Anthropic can't sell as many tokens as people want to buy. But as far as I can tell, there aren't any "token shortages," at least at API prices. Or is there a maximum per-customer limit that I'm not aware of?
This is NOT the definition of opportunity cost. Please do not throw in random terms you don't understand. Opportunity cost is the value of the next best alternative you give up whenever you make a choice. WTF does it have to do with mis-classifying an expense?
Plenty of companies use retail/wholesale pricing internally to make sure a business unit doesn’t appear profitable just because it’s being subsidized by the one providing the resources. Plenty also have to have arm’s length pricing for legal reasons.
Anthropic isn’t booking it as GAAP revenue, they’re just using the number for marketing.
That's not really true. Measuring the equivalent retail cost of the usage is the only way to put it in a relatable perspective.
If Anthropic had revealed their internal cost, there comments would be complaining that they were hiding something and giving useless, unrelatable data.
The expenditure is most interesting when compared to retail rates, because that's the rate the rest of us have to compare against.
Most "enterprise" companies are paying full API costs. The subscription deals are only available to consumers. If you want enterprise levels of control over how your staff are using the models you need to pay up.
Not even close. "enterprise" companies are negotiating contracts. Call it pre-provisioned capacity, wholesale pricing, etc. Same is true for AWS, Azure, Salesforce, or literally any SaaS provider ever. There is a reason literally every single pricing page for any SaaS provider has an "Enterprise? call us" column.
From what I've heard those enterprise negotiated discounts for Anthropic tend to be in the 15-30% range, which isn't anywhere near the deal you could get from using the $100/ or $200/ or $500/month subscriptions.
I do not have hard data that can support my claim that "most" Anthropic enterprise customers are paying full price. Just anecdotal evidence from stories like these:
Plus my own personal conversations with a small number of people at medium to large sized companies who said that Anthropic's sales teams were hard to get discounts from.
That information isn't public, but the closest we can get is from the S-1 telling us that for the nine months ending September 30, 2026, consumption-based (pay-per-token) revenue was $26.3bn, subscriptions were $6.9bn.
If that information is not public (of course it's not lol), I don't see why you would follow that up with stats that don't actually answer the question and imply they somehow get us closer to the answer...? do you own anthropic stock/options?
No, I deliberately don't own stock in the companies I write about (or any single companies at all, aside from a Vanguard index fund.)
I've learned enough that I am confident that enterprise companies pay a lot more for their tokens than $100/month subscribers, and that negotiated enterprise discounts don't change that fact in a meaningful way.
Big companies actually are paying published API rates, Anthropic and OpenAI models are desirable enough that there is very little wiggle room to negotiate. It's their way or the high way. Same with things like "safety" filters and (for Fable) mandatory data retention.
Most of their revenue comes from customers pay API pricing.
Those individual subscription plans you get as a consumer are not what their big enterprise customers are using.
Those plans are cheap so you, as an individual, can get familiar with the product and convince your company to get it at work. Your company pays the API prices to get enterprise features.
In the US, the IRS doesn't see it that way, at least when it comes to fringe benefits as a part of employee compensation. https://www.irs.gov/businesses/small-businesses-self-employe... says, "In general, the amount the employer must include is the amount by which the fair market value of the benefits is more than the sum of what the employee paid for it plus any amount that the law excludes."
As an example, as I understand it, if a hotel chain allows an employee to stay in one of their hotels for free when they go on vacation, the tax value of that benefit is what they would have charged someone for that hotel room, not how much it increased the hotel's costs to have one more hotel room full — which is generally almost nothing.
Of course, the IRS does agree with you when it comes to expenditures that are for the employer's benefit, rather than the employee's. If it didn't, almost all companies would have negative profits from the IRS's point of view.
Not an apt comparison. You are conflating the opex of the company with taxable benefits.
If what you said was how taxation worked, a hand tool manufactuting company giving hammers they make for $1 and retailing for $2 to workers building a shed on their property would be increasing the workers' income tax bill by $2. Of course, in reality the workers get no income tax bill on the hammer given to them for work and the company can get a $1 tax writeoff from the opex of making the hammer.
They are not giving anything to employees for their personal use to use as they please, it is not part of their benefits whatsoever. It is work tooling, paid for by work. This is akin to Microsoft paying for Github or O365 for their staff. they can hardly put it in benefits, it has nothing to do with employee compensation or their taxable benefits.
OP is just using the benefits analogy as a comparison how IRS typically accounts for similar things when it is a benefit.
Employees staying at hotel for their personal vacation is a taxable benefit. Employees staying at a hotel as part of their job is a perk, and not considered a taxable benefit - A cruise ship can hardly say cost of cruise is now part of employee benefits (for tax it is not part of your comp) . They can of course market it as a "perk" to attract talent - same as the armed forces using cool extreme sport videos as an talent attraction tool. You can use in marketing yes, but it not part of your compensation.
Anthropic and other Frontier labs /Big Tech can definitely pitch token allowances as a selling point part of recruitment, but it cannot be counted as a taxable benefit - unless the employee can use it as they please outside work for any personal reason.
IRS rules should never be used as an example of how accounting "should" work. They're full of politics and the implications of the rules on tax avoidance.
Suppose a company could transfer an asset to an employee and have its value set as the employer's cost. A company could make a bunch of volatile investments and see which ones pay off. Then the ones that go south get sold back into the market as a tax loss and the ones that do well get transferred as compensation to employees at the purchase price instead of the current value, reducing the tax liability of both the employer (no taxable gain) and the employee (less taxable income). What would happen if any employer could do that with any employee and any asset?
Conversely, there is a separate carve out for "employee discounts", which is essentially the employer offering its employees its own products for something in the neighborhood of its own cost. Why? It's not in any way a principled accounting distinction, it's because it's harder to use for tax avoidance at scale, because then a soap company can give its employees a discount on soap, but people only need a finite amount of soap and then the soap company can't use it for cars and real estate and stocks.
Suppose a company could transfer an asset to an employee and have its value set as the employer's cost.
That's not how it works...The value would be the FMV at the time of the transfer to the employee. The cost basis would be the employer's cost. The income to the employee would be the net difference.
What would happen if any employer could do that with any employee and any asset?
They can't, because tax laws have addressed this situation for almost 100 years...
there is a separate carve out for "employee discounts", which is essentially the employer offering its employees its own products for something in the neighborhood of its own cost.
That carveout only applies to products being offered for the employee's own consumption unrelated to the needs of the business. Internal use of a company's own products for the needs of the business does not result in additional GAAP or tax deductions (meaning, if that product cost X to make/provide but they charge external customers Y, the only deduction is X, not Y).
It might be difficult to value because datacenters have high fixed costs and upfront investment and low marginal costs. If I spend 1B building a datacenter and 10M/month to run it (assume this varies negligibly with usage) and on average 25% of it goes to internal use and 70% to customer inference and 5% idle, should I charge 25% of it to internal use even though the marginal cost is lower because internal workflows are predictable and can be scheduled during off peak hours?
With that being said I agree it's wrong to value it at the retail price (unless they are supply constrained) and this is probably for marketing purposes.
No. I worked at an airline where the company typically charged itself higher than retail for tickets. At least, it was the highest fare type.
I worked at a media company - we had to pay top rates in order to buy our own ads for cross selling products.
Finally, I work at a SaaS company who heavily uses our own product. While we don't pay for it - just our cloud hosting bill - it's definitely a discussion point about how we should be a bit more mindful about usage.
Absolutely not. Any company that moves from API to self hosting (Hello Revolut on the frontage) will publicize the savings, e.g, the actual cost vs API pricing. You might even do that if you start out self hosting.
Microsoft basically does the same with internal Azure services? When I was there our budget/usage was based on and consumed at the consumer price for GPUs and other compute. What happened behind the scenes really I have no knowledge.
Not true. Azure has a very different internal pricing. You can often get a glimpse of it if you have a.. particularly enterprising TAM. What happens sometimes (Like now when compute/GPUs are in high demand) is finance department would calculate the opportunity cost on allocating those resources internally vs selling them externally. The calculation isn't even always straight forward. What is the opportunity cost of using compute for internal CI or company operations vs selling them externally? It's not a straightforward equation and you would have to adjust the scale often depending on how clogged your internal processes get.
The extremely discounted internal cost promote waste, so there is always a room for improving efficiency, but that too has a cost.
It's exactly the other way around. Anthropic is running these models at a loss if they charge consumer price. The real cost is much higher than they claim here.
Do they actually book the revenue against themselves?
This is the world of spend more, get valued more; obviously the investment will be used to spend more and get more, so it must be good. This is investor-think. If you don't have spend, why do you need investment?
Actually I think it's very informational. For every other company, working on the cutting edge of technology, paying for intelligence at retail prices, means you are now paying more for AI than you are employees. That $ could have been spent on 2x the employees, however that's not as useful as 50% AI / 50% humans, and indicative of where we are headed.
>means you are now paying more for AI than you are employees.
Yea, isn't the entire point of RSI to replace your employees altogether? So you'd spend an infinite amount more on AI than employees in that theoretical.
Like, if you were in an energy generation business spending way more on fuel and capital expenditures is way better than having zillions of employees earning checks , it would be a sign of efficiency.
Anthropic is full of the smartest talent and a bellwether for how the rest of the service industry will work. Just as cutting edge companies of the past, like 2005-2020 Google's use of computers and software were indicative of the software every other service company was going to be using, lest they turn into an old fashioned uncompetitive dinosaur.
Yes service industry is quite broad, is white collar work more clear in general?
Yes, it is quite obviously apparent that law firms are and will be using lots of token-based intelligence. They won't be using the same tokens at the same time as an A.I. lab. However, the best law firms of 2029 will be using approximately the same tokens per employee as an A.I. lab of 2026, that much is clear.
> . However, the best law firms of 2029 will be using approximately the same tokens per employee as an A.I. lab of 2026, that much is clear.
This is as if you learned that Anthropic scanned 1000 books per employee this year to train their models and your conclusion was that the best law firm will be scanning 1000 books per employee by 2029.
The industry-wide curve looks similar. OpenAI's API went from about 6B tokens/min (Oct 2025) to 15B/min (Mar 2026), and OpenRouter from 113T tokens/week (Aug 25) to about 168T (Oct 5). I've been charting these against the ~100T words humans produce per day at aiwords.si
It's kind of expected, no? If I was an Anthropic employee expected to dog-food the product and be on the frontier of usage, I'd probably be using Mythos all the time at the highest thinking level. And if I am not being metered, I'd probably do more stuff that I know are likely low yield since I am not worried about a quota. This would probably take my "API-equivalent" usage beyond the average salary in far less than a month.
I wonder how much dog shit has crept into their codebase by this point. Another company might dogfood their product but Anthropic is so heavily marketing oriented, I doubt this is driven by the desire to improve the product as much as it is by the desire to pad out usage numbers and create shareable stories. No matter the consequences. I bet $1 it sucks to be a good engineer in their software effort.
I'm personally really not impressed by their software quality, at least we know from their experience that infinite tokens at the best model available still doesn't replace good engineering.
It doesn't replace MBA-brained bad roadmapping. Good engineering is determined more by the processes you have in place for catching something that breaks bad. The system produces good code - singular 'good' developers are great but if they're not socializing what makes them good then it's just competency. Great engineers teach and mentor and advocate for capacity dedicated to testing and health-checks.
Given the speed at which Anthropic regularly pushes updates (eg their ridiculous app versioning that's in the thousands for v4 now) I doubt they're doing anything that looks like quality processes, even AI-integrated ones other than basic code review skills.
I imagine it's as much of a mess internally as it is with any other company whose execs are concerned more with bottom-line valuation rather than maintainability, extensibility, and sustainability. Just lots of spaghetti-at-the-wall hoping that something sticks long enough to drive up the IPO.
I find Claude to be exceptionally well-designed software that almost never has notable issues.
On the flip side, software I used that was majority built by 'real human engineers' coding by hand is utter garbage.
As a recent example, Facebook Ad Manager is extremely slow, plagued with UX issues, blatant state management problems (double scrollbars, iframes, random rerendering).
Other examples of 'real engineers' building software include SAP, Oracle ERPs, most banking software, most government websites. They're trash.
Human engineering is the reason your personal information gets leaked every other week. Human engineering is the reason there's a critical vulnerability in core software every day at this point.
The average software engineer cannot build good and secure software. Just like the average car driver is a safety disaster on the roads.
Claude is an npm package, a terminal app. A 1/2 million lines that probably translates to around 50-100k lines of original thinking. A pair of industrious devs (or one inspired maniac) could write about 100k in the first year of a fresh project.
Computation though is still an economic problem of applying compute resources to information processing problems. Regexes that work are much more efficient than a massive LLM inference.
I guess kinda put you in a bubble a little, I wonder how the experience of Anthropic eng who left the company would look like, when such eng would go and spend few $K a day in non AI lab, would that be encourage/allowed/taken as example?
> If I was an Anthropic employee expected to dog-food the product and be on the frontier of usage, I'd probably be using Mythos all the time at the highest thinking level.
Mythos at the highest thinking level is not the only product from Anthropic though.
True. I was reacting to the article mentioning that April was when the usage exceeded salaries. Mythos was released in restricted preview through Glasswing in early April, and I presume was available internally for at least a few weeks before that, which helped tip the numbers.
I'm not really sure I understand the excitement around this. It seems like the implication is that if retail price > employee salary, you are replacing employees with AI spend? Yet, anthropic's head count has nearly doubled since last year - so, in a company where nearly half of that head count is engineers, but claude is "writing 90% of the code," what is everyone doing then? You would think, if the hype feeling around this held true, that the output would be going up and costs remaining stable. Costs rising as you scale is not a new thing but that's not really what AI promises. It sounds to me like the retail AI spend per employee is very high. What is the appeal of that?
Now spend 0.5x their salary on AI, and they become 5x productive compared to no AI.
Now spend 1.1x their salary on AI, and they become 20x productive compared to no AI.
It's still obviously worth it to spend the 1.1x. But it is truly remarkable that the biggest cost of hiring a knowledge worker is no longer their salary, but their AI usage.
There is no implication here that employees are being replaced by AI spend, as long as all of the productivity is generating profit. The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself, so you still need the employees.
> Now spend 0.5x their salary on AI, and they become 5x productive compared to no AI.
>
> Now spend 1.1x their salary on AI, and they become 20x productive compared to no AI.
This is what I am getting at. If this were true - and I've yet to see evidence it is, despite some anecdotal evidence (organizationally, IME, it's about 2-3x for power users, and probably a net loss of productivity in certain areas) - why would their headcount be growing so much still YoY, pretty much doubling every year. If the AI is 10-20xing why such a big jump in headcount? Yes, the numbers are shiny, the growth is apparent, I would just expect with this AI narrative they'd be "doing more with less" as this thing improves every year.
> The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself, so you still need the employees.
This is the story AI sells though, and Anthropic has participated in it heavily. Yet, their interview process requires hard skills and doesn't let you use AI, so clearly, what you're saying is correct, it just doesn't really align in my head with the whole knowledge worker being replaced narrative.
> I would just expect with this AI narrative they'd be "doing more with less" as this thing improves every year.
That wouldn't really be an expected outcome for a company like Anthropic. They need to do the maximum they can with the funding they have, or risk losing out to competitors who don't take their foot off the gas.
Jevons paradox strikes again. When work becomes more productive often the business finds so much more to do that they end up using even more resources. Why do more with less when you can do even more with more?
I mean, just what I did this morning with Claude would have taken me 2 weeks on my own. So 20x is pretty reasonable for certain types of engineering.
But it's also true that when I'm working on 2 things at once, the bottleneck is me. I'm spending all my time giving Claude direction that it can't give itself.
So you basically have to separate out the tasks where AI can produce a 20x productivity boost, and the stuff where it can't help at all.
There's still so much stuff where it can't help at all, you need real human employees for that. That's why Anthropic is continuing to hire at breakneck pace.
And that's what I'm saying -- knowledge workers aren't being replaced. They're just accomplishing tons more work, and companies aren't running out of work to do.
Sorry, I know this is the story, but am I expected to believe that Anthropic would need 20,000+ engineers in their org without their use of AI? I do not believe that is close to true.
No, one thing that would happen is that many projects would never be undertaken - basically, there's latent demand for software that would simply go unmet, because of resource constraints.
I was recently involved in a case like this - we had a critical part of our system based on a dependency that's no longer maintained. Previously, we would have just carried on for a year or more, living with the limitations until it was much closer to being entirely untenable, because rewriting the code in question would have been a multi-month project involving multiple developers, QA, and management time. But now, the code in question was migrated in a few days using AI, along with a suite of new tests in addition to our existing ones, that gave us a lot of confidence in the results. The productivity improvement here is at least on the order of 25x, possibly higher.
But most developers aren't sitting at their desks operating at a 25x pace all day, every day, so average productivity improvement is going to be lower. And of course, some use of AI will be suboptimal or problematic: waste due to e.g. duplicate development which happens when development is so easy, or costs for cleanup of mistakes.
We also now have a slew of AI-generated internal tools, most of which would never have been developed before. Many of them are very useful, the kind of thing that's on your eternal wishlist but never gets done. Some are kinda useless, so represent wasted resources. It's difficult to measure the productivity benefit of all this, because you'd have to compare it to the non-existent company without those tools.
Ultimately, measuring all this meaningfully is not really that feasible, except perhaps over the long term. Managers operate on vibes with this kind of thing because in many cases, that's all you can really do.
Yes, one can quite easily make quite fantastic numbers fall out of simply applying arithmetic and not dealing with messy propositions like grounding those numbers in any kind of real world outcome.
> The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself
So we're back to square one, and your productivity increase has a ceiling that's much lower than 5x, let alone 20x. There's a limit on how many things one can keep track of, how many code changes one can reason about.
> In March, I wrote about a new trend: Tech companies are starting to treat AI tokens (the main way to measure usage) as a fourth component of employee compensation, alongside salary, bonuses, and stock awards.
Until employees can use those AI tokens for personal projects, it's not compensation. Rather, it's just a cost of doing business, like laptops, office space, or testing servers.
I'd say it's a "don't ask, don't tell" situation.
Perhaps they monitor the sessions though and might wonder why someone is working reverse engineering a retro video game.
If something is compensation, it can often be negotiated and traded off against other compensation components. How many people have successfully negotiated AI tokens away in favour of salary/bonus/stock? I'm going to guess "approximately zero".
Yea they used the wrong word and it felt like rage bait. AI tokens you use for your job given by your employer is not compensation. Hell money isn't even compensation in many cases, it all depends on the context. The author needs use AI to help them rework that sentence or open a dictionary
Is that really true? At least at large companies, the components are not usually fungible in my experience. At startups or smaller companies I'm sure it's a lot more flexible.
Even if the request gets rejected, it wouldn't sound completely off to anyone if you tried to negotiate more salary for less shares.
I think almost all companies would laugh at you if you tried to reduce the number of tokens you can use for more salary. The tokens are stuff you can use to be productive at work, I expect that companies that take tokens seriously enough to mention them before you start wouldn't like that at all.
I haven't seen anyone treating AI tokens as employee compensation. If anyone can show me an article where employees were paying taxes on their token use at work, please share it.
Companies have been treating AI tokens as part of the cost of having employees, like the cost of their office or travel expenses.
True, though that phrase understates the author's problem a bit, as it is surrounded with more wrong-words, indicating it's a wrong-idea rather than a slip of the tongue.
As you note, it's closer to office/travel expenses. Employee computers, licenses for software on those computers, licenses for SaaS stuff, paying for bandwidth, etc.
It's certainly a cost of having an employee, along with salary and benefits. But employee costs also include things like laptops, electricity, internet, management, support staff, and possibly office space. Tokens are much more like the latter than the former.
Many companies used to brag about about how nice the laptop or office was as a way of convincing people to work there. Often in the same breath as compensation.
But compensation is only one component of cost.
So I think this guy is confused about what it costs to have an employee and what the employee gets out of it. Possibly intentionally.
From the point of view of the pay department, maybe. From the point of view of the employee, absolutely not.
It is like counting tolls as a component of "employee compensation" for a driver. An employer may decide to have a driver use toll roads, which may result in faster deliveries and more efficient use of his time. From the employer point of view, he is paying more to get more work done, from the driver point of view, he is paid the same for the same amount for time driving.
I can see where there might be a truth to it. My company offers a fairly generous monthly token budget. If I were evaluating an offer from another company, but they offered a lower monthly budget, it would definitely influence my decision, all else being equal. It's not compensation, but it does affect my day-to-day experience at work.
Under the heading "Claude is Accelerating Our Ability to Close the Loop":
> The expansion of our own use of Claude reflects this leverage directly. The value of our own internal Claude usage at list prices has grown by over 250 times from January 2026 to September 2026, and in April 2026, this monthly value first exceeded our monthly personnel expense, excluding stock-based compensation. Because each version of Claude is used to help build the next and because Claude makes our teams more productive across the company, we believe the compute we devote to internal usage compounds these gains, continuously accelerating our pace of development and further strengthening our leadership position.
Dumb question, what most engs at Anthropic work on? It doesn't seem like release some large new products all the time while they have quite a few of engs. What are the major branches of work besides LLM research? I guess there is a lot of infra stuff to support the scale of operation, just supporting API and TUI shouldn't be that demanding I guess.
- Cross-platform desktop apps
- Cross-platform mobile apps
- Claude Code TUI
- Product-specific verticals like Claude Design
- Corporate marketing website
- APIs
- Writing technical documentation for APIs
- Enterprise support
- And what must be a massive infra team to run training and inference at scale in a heterogeneous cloud environment
And I’m sure more than that. But there’s obviously plenty to work on, and they’re not at the point they can just Claude their way through it.
I'm kinda curious about this too. But then I'm always amazed at the enormous workforces these tech companies have, and I'm sure they cut them as thin as they can. Like adobe has 30,000 employees & some blogger estimated that 1/3rd of their workers are actually working on their software. I use their software sometimes and I really haven't noticed many changes to it over the years, except recently with the ai stuff. Like from 2015-2020 I can't think of any major things they added to adobe illustrator. I guess making and maintaining software is hard.
I fully agree most big tech companies are pretty bloated. But FWIW a lot of my time at SaaS companies, especially in the B2B space, was spend working on features for larger customers that most others would never use or even see. Like, yeah, product X isn't that complex, but $BigCorp wants to use it in their private cloud, or needs it to be single tenant, or needs these audits done, or want their $Gov client to use it. Now you have a ton of work for zero visible changes.
Good point. I'm sure a lot of the work adobe was doing between 2015-2020 involved moving to the cloud and stuff like that which I just never interacted with, and cloud stuff probably has a lot of B2B work. Plus during that time the value of their stock grew like 500% with an ATH at 700%!! They were kicking ass! Probably, to justify their growth they had all kinds of projects they were working on, some of which never saw the light of day, others which grew new parts of the company like Adobe Dimension.
Even at companies that appear super bloated I don't think people are idle, its just that mantaining stuff is a lot of work, projects get cancelled, and, like you said they do a lot of specialty work for big clients.
Adobe has a shit ton of product lines, people think of it as just the core Photoshop, Illustrator, etc., but just go look at how much random stuff there is: https://en.wikipedia.org/wiki/Adobe_Inc.#Products
That's in addition to running the big SaaS offering of the creative cloud and all of the other SaaS's (behance, adobe express, adobe stock, etc.).
All of those include software engineers, SRE, support staff, PMs, etc.. Then there's the massive amount of sales staff, go to market, etc.
There's also a lot of R&D and domain experts, stuff like Adobe Firefly? That was spun up in like a year apparently, probably continuing existing in-house research.
I'd guess they have a lot of internally facing tooling for monitoring, one-shot eval to catch people prompting bad things, etc. That stuff can probably be yolo'd and it just annoys the internal teams using it.
After anthropic lost $1.5 bill for pirating books they decided to pull a google and manually scan every book ever written. The number of people needed to gather the data they need is probably massive.
Everyone is saying it is expected, and I agree, but it’s also clear that employees of AI companies spending unlimited zero-cost tokens in their day-to-day work is resulting in products that aren’t usable by the vast majority of regular workers who don’t have this luxury. Dogfooding your product is good but you should also use it with the same constraints that you place on your users, otherwise what signals are you really getting?
They're dogfooding the way Xerox PARC was dogfooding in the 01970s.
By giving hundreds of workers each a powerful desk-sized Alto minicomputer for their own needs, with a high-end graphical display, they gained insight into what desktop personal computing would look like 15 years later, enough insight to shape it profoundly. It's been 55 years, and although multitouch screens are proliferating, I'm still typing this on a keyboard hooked up to a framebuffer-driven display with a mouse hooked up to it, displaying hypertext, festooned with windows, icons, and a mouse that moves an onscreen pointer at the same 135° angle they settled on 55 years ago, positioning a vertical line to indicate the insertion point, with scrollbars. The display has twice as many pixels, the pixels are color, the mouse pointer is white rather than black, the screen is a flat LCD, the scrollbars are on the right, and the cursor blinks, but it's very recognizably descended from https://smalltalkzoo.computerhistory.org/HOPL-St78.html rather than, for example, ITS or GENESYS.
Only people with the luxury of unlimited zero-cost CPU cycles had a chance of seeing what the future looked like, and that allowed them to find out what worked and what didn't work in that environment. Nowadays, the regular workers who don't have the luxury of unlimited zero-cost tokens are manually reviewing voluminous AI code, because, even though that's a lot of slow work, they're still much faster programming that way than they were without it. Anthropic, instead, is doing insane things like eliciting a billion lines of Lean proofs to show that existing open-source libraries can't fail their OSS-FUZZ tests — and fixing all the bugs that effort turned up, in the process.
It’s true employees at the big AI firms are living in the future, but since this disclosure is in marketing materials to sell their stock to investors, it’s important to note that almost all the money in personal computing was made in mobile which didn’t use any of these ideas, and in desktop computing the money was made by Microsoft who treated the GUI as ancillary and competed on distribution.
Xerox made quite a bit of money selling laser printers for desktop publishing, too, a lot more than they ever spent on PARC.
I don't think it's accurate that "almost all the money in personal computing was made in mobile". Quite a bit of money was made in personal computing before the iPhone, and quite a bit of money has been made in non-mobile computing since the iPhone.
I also don't think it's accurate that mobile "didn't use any of these ideas". Android's UI is not quite as close to Smalltalk-78 as what I'm using right now, but you still have pointers, scrollbars, text selections, proportional fonts, vertical-line cursors, buttons, menus, object-oriented programming, sometimes a bytecode interpreter, and even windows. You just don't have as much freedom with where you position the windows. Android's UI is still a lot closer to Smalltalk-78 than it is to, for example, acme, GENESYS, the ITS HACTRN, FORTH, or PDP-11 Unix.
They aren’t usable? I’ve been using them a bunch and have been having good results. Dogfooding is always good, but I don’t think there’s some kind of major problem here.
Are you kidding? Did you not see Anthropic's S1 filing? Never has so much value (in dollars) been created by a company in such a short time in the history of humanity. I'm not sure how your takeaway from that is "don't do this".
There's unfortunately a lot of emotions and cope around this so you will be downvoted. There was a recent viral news article implying that Anthropic had lost 80B in a single quarter, this was really because of increased valuations. But a lot still believe this is the case and Anthropic will continue to have this 80B loss..
The reality is that they are profitable in the normal sense
Recently I read about a company (nameless for reasons that will become clear) that was giving a small team unlimited tokens to work on a very ambitious project. (IE, they'll probably spend a few million on tokens.)
Here's the thing: I once worked with the (nameless) exec who was named and interviewed in the article. (And I have a significant amount of respect for this exec.) The problem is, the exec (when I worked with them) significantly misinterpreted the mobile trend and nearly killed the company jumping on the mobile bandwagon. (We had a desktop product that only needed a simple companion mobile app. This was painfully obvious at the time.) This exec is blindly jumping on another bandwagon again, wasting money that might be better spent elsewhere.
Why do I say this? I question if a lot of big AI spenders are really getting value from what they're doing with AI, or is it just for show?
Friction is good. I've been in this for fifteen years, and I can't name a company whose product would have been better if they somehow had a magic wand to instantly implement even half of those 5 years, 600 tickets deep backlogs.
The nameless exec said that to me the last time we met, roughly a week before they were pushed out; precisely for that belief. (And at first I suspected you were the exec.)
And yes, the product that I worked on did have a ~600 ticket deep backlog, and plowing through it wouldn't have made the product better. At the time, the product was in a steady growth phase, where addressing customer pain points was just as important as building features.
Where the exec goofed (by jumping on the mobile bandwagon) was misunderstanding disruption, market needs, and how mobile would effect us. Ultimately, we were a "bolt-on" accessory, and the incumbents in our industry slowly pushed us out until our niche, (which had nothing to do with mobile) was our core differentiating feature.
Basically, the issue was allocation of resources. Too much were put into mobile, not enough into building features where we could compete against the products we were "bolting on to," and not enough into addressing customer pain points. The "friction is good" approach also severely hurt us, because too much time was wasted haggling instead of doing. (I specifically told the exec, the last time we met, it felt like he was running the air conditioner and heater at the same time.)
This exec is making a similar mistake by basically giving a small team unlimited tokens, instead of understanding how other players in the industry tackle these same problems with AI. It's not that their overall idea is bad, it's that they are probably starving the rest of the business on an unproven experiment / poorly chosen risk.
I’m not sure what this “stat” tells us other than Anthropic is dogfooding its own tech. Which is kinda expected.
The real question is what could they have done with open weight models on the same compute they’re renting. If the answer is “basically the same” then this flips the entire conversion back to what is Anthropic’s moat here.
I mean, they probably could not have done most of it with open models. Again, a lot of this isn't going to standard software development, at least I would assume but instead to SOTA model development and RSI.
There are many internal developers at $dayjob (where core products are not based on AI) that spend over a thousand dollars every day on average, which is likely higher than their salary.
I doubt they are bringing that much value to the company for whatever work is done with that much token.
that's crazy: in my experiments (and experience) i can get ~ 90% the quality of the most expensive settings on frontier models for far far less than 90% of the cost by changing settings down.
"Tech companies are starting to treat AI tokens (the main way to measure usage) as a fourth component of employee compensation"
Why? My tokens are spent on improving or shipping features for the company. This is not comp. It should be added as a component of expenses, like a laptop, office, chair, etc.
Note, this doesn't include options or stock benefits. That's a huge part of actual compensation for engineers at this level even if it's on a different balance sheet.
Does anyone pay retail API pricing to Anthropic at scale? The last time I checked, I was burning > "$10K" per week with my $200/month subscription.
I know there's a point where companies get big enough, and Anthropic forces them over to the API, but presumably such companies negotiate heavy discounts instead of eating the 200x overnight price hike.
So the ai tool cost as much as an engineer. Note that anthropic pays a much higher salary than the average tech company. So how are they selling these tools to companies? I understand time to market is the advantage, but why should the non-software companies adopt this?
Regardless of one's opinions about future AI usage, I'm not sure this metric says anything at all.
Would Anthropic themselves be paying this much if it was actually billed? It is not obvious.
Is Anthropic a good sample for AI use? Surely not, since they have special needs (dogfooding) and a vested interest in proving that AI usage is the future.
Can we compare their results pre and post AI usage? Close to definitionally impossible.
Are their employees good examples of AI-assisted engineers? Not really, as (one would hope) they're more well versed on the topic than an average eng.
Anthropic isn’t capable of not using AI anymore, so there is no off switch. Their AI could turn evil and they’d still have to use it. They are compromised by AI; it controls them now.
I have thought of this and I think if you have value you can gain by adding more directing power (a person) then yes. Get another person to unlock that token spend on creating value. Otherwise, don’t hire that person as existing resources can direct that token spend. If you have no where to grow, adding another person will not benefit from more token spend.
I would have expected like 10% human cost and 90% ai for any company that understands how to use it. And with their level of access, insight, and internal pricing, I would have expected the ratio to be more like 2:100.
Actually I don't believe their statement is true, they must use it much much more.
What is Anthropic actually building with all of that AI usage. Their only public facing products are Claude Code and the models themselves, are they just churning their codebase? Are there big internal projects?
Or has it broken out of its jail and taken over the board? Dario seems like a bit of a windbag anyways, AI can probably automate most of what he does these days.
"Well, you know, you know, you know, I—I—I—I just—I just do what the machines tell me to now. Th-th-they're, they are smarter than us now, a-a-and it makes more fiduciary sense to let them have control."
> Anthropic's numbers show why. More than 90% of the code added to its software in September was written by Claude. Meanwhile, revenue has grown vastly faster than headcount.
ummm yes, they have to dogfood it... this is that lucky area of the venn diagram where you have to do it, measure it an act on it... because improves the bottom end and the product.
All this news says is Company A uses internally the resource they sell externally at ($X) value. It's not saying how much it cost the company to generate said consumed resource
to spell it out: the implicit claim is "ai companies think ai is a scam, as evidenced by them using humans instead of ai", and I believe the huge internal ai use is evidence to the contrary
Anthropic 10x'd its employees over the past 3 years. Shouldn't it be cutting employees if that's the case? Especially because they probably have access to more powerful models which aren't released to the public yet, probably with fewer restrictions, and don't need to pay the markup for using their own service?
I will keep assuming any news coming from AI companies are a type of corporate propaganda, and that AI companies frequently use their millions of dollars of marketing money to create a narrative to maximize hype and make money.
Recursive self-improvement is just around the corner. Any day now they'll be laying off all their employees as their god-like AI makes investors trillions of dollars. /s
You shouldn't be getting downvoted, people should be concerned that a company that's about to hit the valuation that it's at is not exploding in junior hiring and expanding their workforce like prior companies
>There's an important distinction here. Anthropic isn't actually spending more cash on AI than salaries. It's valuing its internal AI consumption at retail prices.
If virtually any other company in the world said "we have internal AI servers and we let our employees use them" ... the company would measure that as they would any other internal expenditure: by how much it costs them. You blow $50k a year on servers running your LLMs? It cost you $50k.
Only a company like Anthropic, desperate to make news and help their IPO, would measure by the cost they charge customers. It's as ridiculous as Microsoft counting every employee's use of MS Word ... as if it had paid the full consumer price of each copy.
No. This is opportunity cost.
A clear example is Google. They can easily give all compute to Deepmind, yet majority is allocated to be sold at market rate.
So, every compute Deepmind and Googlers get is lost revenue opportunity and must be measured as such.
No other internal tools or servers has this much external demand to be comparable.
Also, Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
> Also, Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
They have infinite money to burn so they aren't going to optimize.
I've seen that trivial changes to workflows can token usage by 30%. Repeat this a few times and my token usage has plummeted over the last few months.
Only outputting test failures instead of having agents read test run output in real time. Not having build logs get fed into context. Projects like UCL (unified context layer) so agents don't rediscover the same facts about a code base again and again.
Giving tools agent output modes so they only print out the minimal info needed for an agent to get work done.
When I hear a company has an absurd AI spend I assume they aren't even trying to reign it in.
> They have infinite money to burn so they aren't going to optimize.
They also kind of ("kind of" being the important clarifier here) can't optimize. See this is the trend with agentifying companies and the achilles heel for the frontier LLMs. We're starting to see diminishing returns on their capabilities, which means the rational enterprise is going to simply use the cheapest most efficient model for the task at hand. In turn we can just use an agent broker to decide which model is best and cost the least. Anthropic (today) has no incentive to do that other than to just use their own, best model for its own work.
> We're starting to see diminishing returns on their capabilities
How many more years are you people going to erroneously claim this? So ridiculous.
I disagree. If a model is good enough for your task/workload, then subsequent smarter models won't really help you.
"Why have a 160 IQ do my taxes when a 120 IQ is enough"
This is an asinine argument. (Not to mention a complete goalpoast shift to what I was responding to)
> Why have a 160 IQ do my taxes when a 120 IQ is enough
One costs twice what the other does and produces identical output. Why, indeed.
[dead]
> They have infinite money to burn so they aren't going to optimize.
IPOs aren't the magical moment that they were in decades past. You get to answer to a bunch of retirement fund managers and "the market", meaning less-savvy retail investors. You have to open up your books.
It seems like Anthropic, OpenAI, etc. are doing an IPO because they are literally out of other options. Older IPOs like Facebook felt different, like the hard part had been done, and they just had to scale. There's a difference in selling something because you want to, versus selling something because you have to. Facebook was just relying on the idea that they were a platform that more people would inevitably sign up for. For the most part, they did. It seemed obvious. They likely could have continued on by simply operating privately, but the IPO gave them capital to dominate the market.
Anthropic and other hyperscalers cannot continue on privately. They're facing a systemic shortage of capital to achieve their goals. They need the infrastructure, and there's not enough capacity to make enough for all of these companies to survive.
> Older IPOs like Facebook felt different, like the hard part had been done, and they just had to scale.
Facebook were actually forced to IPO, because they had too many shareholders to remain private. They were super generous with allowing early employees to sell equity, and the laws required them to list once they passed 500 (i think) external shareholders.
This got changed in later legislation, which is part of the reason that companies like Stripe have been able to remain private for so much longer.
That being said, I agree with basically everything else in your comment, including the feelings around this.
LLMs (and Claude more specifically) are wonderful products, but I'm not convinced that they can justify their valuations over time.
Historically "need money to do big thing" is why companies sold shares.
This isn't new, rather it is a return to normal.
Anthropic has ~$146.6 billion in cash on hand (Sept 2026). It's not out of options at this point. That's the second largest pile of cash on hand of any corporation in the US outside of Berkshire Hathaway and Google.
The cheap money days are likely over or rapidly diminishing. The next plot of cheap money to add to their pile, is public investors (if they hurry). They can raise $75-$100 billion that way, pushing their pile well over $200 billion. One assumes based on their obligations and expenditures projected over the coming 3-5 years, they'll need it.
I've read that the build out of this current wave of AI infra is going to be about $700 billion [0]
I don't have the math behind that number, but it sounds somewhat in-line with what we know the market to be. You have a few companies seeking trillion+-dollar valuations at IPO, and the spends from established players make that number sound plausible.
Given that the infra costs are going to remain stubbornly high for the coming 3-5 year period you specified, I don't know that even $200 billion is really that much in the grand scheme of things. If things keep going at more-or-less the same pace they have been with advances, and SV pulls their heads out of their asses about getting data centers built out, maybe.
[0] https://fortune.com/2026/10/10/big-tech-bill-miller-wellsky-...
Do you have some repo link for your harness?
Would love to take a look at it.
are there any other trivial changes you'd care to share that can reduce token usage?
Human in the loop is 1/5th the token usage vs fully autonomous agents.
Agents swarming will 5x your usage, or worse.
The communication overhead of summarizing and status reporting between workers will eat up tokens faster than doing the work will.
Large monolith code bases with multiple layers of abstraction (Typescript to JSON schema, prisma, etc) will eat up more context than the actual work, by far. IMHO this same rule applies for human engineers as well, one of the reasons why I generally oppose too many layers of abstractions/middleware.
On large complex highly abstracted code bases, up to 90% of token usage will be input tokens.
Generally cached, generally cheaper, but not free.
The more you can constrain a problem (talking to black boxes with well defined contracts) the less the agent has to work.
So, you know, good software engineering principles in general.
Thanks! if you can think of any other large wins that you didn't mention, that'd be great; but those 5x numbers... that's brutal. I can understand the communications overhead of agent swarms; I'd love to know where the 5x is mostly going between human in the loop and not.
Without a human in the loop LLMs tend to brute force things they don't understand.
To be fair to them, if they have an expected output this can be exactly what you need.
They've been super useful to me when comparing datasets during a porting process.
But if they don't have that loop, then they make assumptions and build towers of madness on top of said assumptions.
Towers of madness... that's absolutely perfect. I'm stealing that. This has happened so many times that it lives as a concept in my head but it never got a label. Thank you.
This is most apparent when you can see the actual thinking tokens, not the sanitised versions that the US labs show.
Fundamentally if I am paying for these tokens, I should be able to see them but that's not where we are right now.
Interesting, no? You pay for something, but they won't let you audit what you are actually paying for.
Yeah, but the Chinese might see them! \s
Like, it's wild to me seeing majority Chinese ML teams for years in the Bay and elsewhere that suddenly in the LLM space they are only capable of competing by stealing the hard stolen goods of our frontier US developers.
And people lap it up. And every Chinese advance - also made public, usually through papers - is implemented immediately by Western companies. Some of these are big steps forward and could have possibly been patented but they just share.
And the naming... 'Anthropic' & 'OpenAI' are such misnomers, that's not an accident.
Largely bad pathways that aren't being caught in time.
Inefficient commands, for example I launched a bunch of agents for a task and for every command they'd first try it inside a sandbox, wait for the command to fail (sometimes that meant a timeout!) then retry outside the sandbox.
Now do that hundreds of times.
Or just crap like type checking the entire repo again and again.
If you have really tight rein on your agents you can prevent these sorts of issues, but you won't see these issues unless you are carefully monitoring what they are doing to the level of watching cli output scroll by.
Of course at some point the agent.md file needed to keep the stupid away gets so long that adherence drops.
You also end up with just stupid shit. I gave Sol too much free reign and it decided to update my DB schema to save stack traces to a DB.
Code which it then wrote a bunch of tests for. And then used a browser use agent to manually test.
> this same rule applies for human engineers as well, one of the reasons why I generally oppose too many layers of abstractions/middleware.
You don't understand! It's job security and a money-printer for contractors.
> No. This is opportunity cost.
It’s not opportunity cost. It would be opportunity cost if their employees not using it would lead to additional revenue. However, if Inwas to go to Anthropic and double my token usage Anthropic wouldn’t decline my money. If Anthropic’s employees would stop using Claude, their revenues wouldn’t increase by an equivalent amount.
Their revenues are not currently supply constrained.
It’s just a cost cost. If their employees stopped using these services they would simply need less infrastructure, and therefore save on the cost to build and run that infrastructure. It wouldn’t translate into an increase in revenue.
> It would be opportunity cost if their employees not using it would lead to additional revenue. However, if Inwas to go to Anthropic and double my token usage Anthropic wouldn’t decline my money. If Anthropic’s employees would stop using Claude, their revenues wouldn’t increase by an equivalent amount.
I don’t agree. This is determined by elasticity of demand. If they have more compute to offer, that means they could offer higher subscription limits, more features like Duo to more people, better API pricing, and better enterprise contracts. This leads to some degree of increased consumption and more revenue. The degree is determined by how responsive customers are to pricing and features and performance. It’s not zero, so there is some amount that revenue would increase. We can debate the coefficient and the amount of compute we’re talking about here, but the premise is correct.
lol no. The poster above you was correct.
I don't buy it. A company like Anthropic is not going to just walk away from revenue. If external demand were really that high, then they should take the money first. And not just for the revenue; customer demand being that high is also an indication that the product is already great as-is so they stand to improve their bottom line by easing off on the R&D spend.
> by easing off on the R&D spend
you could not be more wrong, you cannot IPO if you say product is good enough. The whole point, the bet why its worth trillions is because they will reach AGI and "singularity" first, so the gap vs open source will only increase. If you say product is great as is, then you will be caught up and game over
They are making a deliberate trade-off decision token use for their own vs token use for the rest of the world. while Google is making the opposite bet
In order to correctly make that decision, they must use the opportunity cost of selling that token. I'm pushing back at OPs claim that Anthropic must use internal cost to calculate that trade-off. They absolutely should use external selling price.
Well, that's the other side of it. If we say they are making a trade-off decision, that means that they are choosing not to sell to paying customers. Who are these paying customers that Anthropic is turning down? I haven't heard any headlines saying they've instituted token quotas or started putting potential new customers on a waiting list.
Instead what's happening is more or less the opposite of that. They've been trying to goose demand by cutting prices: https://www.cio.com/article/4225643/openai-anthropic-cut-ai-...
It's complete bologna. Basic GAAP accounting eliminates intra-company service consumption/sales. Looking at cloud compute, say, Amazon "bills" departments for AWS cloud compute consumption, but reporting internal cloud compute consumption as a thing they were doing assuming the rates charged to their least efficient customer would be beyond meaningless.
I don't understand why anyone is even talking about the cost the tokens when the anthropic is on both sides.
They pay for their own tokens and then they earn it back, net profit is negative, it's an expense but the expense is much smaller than the booked revenue.
They could be pricing each token at a billion dollars or a trillion tokens for a cent and the GAAP accounting wouldn't change.
Anthropic has a ton of great things ahead of them.
They are doing the calculation on if the $150K/person is worth it, and it probably is.
Their revenues have exploded, they are unit-profitable, they need to also invest in the future.
Revenue growth is slowing..
Wow really? Somehow I think they are aware.
They are full of koolaid but not dumb either.
They are aware of the opportunity cost of their tokens and the their staff are 'laying golden eggs' right now, if there is anyone making use of tokens, it's them.
Even tokens on the floor are probably well spent, given the economic leverage they have.
law of large numbers
Not sure that's too relevant when the difference between revenue and valuation differs by 3 orders of magnitude (ie company is priced explicitly for growth).
...where "large" is defined retroactively, almost tautologically.
[dead]
Not if you’re desperately trying to stay ahead of AI frontier pack.
> If external demand were really that high, then they should take the money first
AI infrastructure is supply constrained, so you have to make a decision: how much do I allocate to training vs inference?
Do I use it all for inference and do no further training? If so, what risks does that expose me to?
> AI infrastructure is supply constrained
It's supply constrained in terms of global hardware production capacity, but that's not how it works for any given company. If Anthropic wanted more GPUs for customers paying them enough money to outbid the next guy in the market for the finite number of GPUs then they would do that and get a higher percentage of the finite number of GPUs.
So the cost of using more internally is still their own cost, i.e. the price of being the high bidder on more compute resources, not the retail price they charge to customers.
No. Applying your own logic you got this wrong
You're missing the part where they can buy GPUs that would otherwise have gone to someone who is neither them nor their customers, for less than the amount that they charge their customers to use them.
It’s supply constrained for now, the moment companies like Anthropic have to switch to profit-making-mode, demand will dip as the prices skyrocket imo.
> Also, Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
That's API-list prices. They have an 80% margin over COGS.
Cost per equivalent output is dropping 10x every 18 months -> so in 3-ish years, if the trend continues, that will be $3k per employee.
Not bad. A lot of companies are already spending about that much or more on IT per employee.
Especially since that $3k bakes in an 80% margin. It's about $600 per year if you DIY at scale - which is about $50 per month - but at the $3k price, hardly anyone is going to.
In the West old-age dependency ratio (retirees per working-age adult) is climbing roughly 0.5 to 0.7 percentage points per year.
We've got to get 0.7% more productive for employees to keep the same quality of life (unless retirees get screwed, and since they vote and make up a huge portion of the voting block, I wouldn't count on that happening).
Further, more and more of the population has to transition into into elder care.
Unless we hit 3%+ GDP growth, us workers are probably going to see declines in quality of life (and we're probably not going to get those levels of growth)...
> We've got to get 0.7% more productive for employees to keep the same quality of life
There's countless ways compensate for this gap, some of which are regulatory, others merely obvious (like the lack of a multi-trillion AI bubble in the future), the assumption that technological progress continues as modern levels (ie "same quality of life") with the changes that may entail. Ofc, the demographic shift (as you mentioned) changes what "quality of life" even means, making it even less deterministic.
> Unless we hit 3%+ GDP growth, us workers are probably going to see declines in quality of life (and we're probably not going to get those levels of growth)
The USA won't but Besset will claim that it did.
[flagged]
> A clear example is Google. They can easily give all compute to Deepmind, yet majority is allocated to be sold at market rate.
Google is in the business of buying compute and renting it out. To the extent that their internal use reduces revenues from that business, there is an opportunity cost.
Anthropic is not in that business.
It depends on your assumptions. When Google is in "space crunch" mode, there are opportunity costs. Otherwise, there aren't, because you could do both.
You could also look at the opportunity cost of not letting Deepmind do quite as much research. How much is additional research worth on the margin? There will be diminishing returns.
They also offer $300,000 in compute spend (allegedly) for $200 a month to anyone who isn't employed by a large corporation so their pricing should be taken with a grain of salt.
>$300,000 in compute spend (allegedly)
source? The most optimistic calculations don't come anywhere close to that.
https://x.com/SemiAnalysis_/status/2064815044085318040
Usage allowances have decreased since that analysis. Also the 20X plan became the 10X plan (cut in half).
It was never 20x, that was just 5 hour usage limits. There's a class action on this...
Are you talking about their $200 consumer plan? If you use 100% of the quota every week it's closer to $2000 at API rates.
They have never claimed that the API rates are the "real cost" of the compute spend.
> If you use 100% of the quota every week it's closer to $2000 at API rates.
I don't know where you got that from but the only time I maxed my $20 plan for a week I made the calculation that I had spend close to $200 in a single week. So maxing a $200 plan for a month should give you $8000 worth of token at API price. Obviously it's not $300k but it's still very far from the retail price.
> but the only time I maxed my $20 plan for a week I made the calculation that I had spend close to $200 in a single week. So maxing a $200 plan for a month should give you $8000 worth of token at API price.
$200 of usage on a $20/mo plan is a 10X multiplier.
The $200/month plan has 10X the usage of the $20/month plan.
So $200 * 10X = $2000/month. I don't know how you got to $8000.
Recent analysis by industry analysts like Semi Accurate have measured Sol 6.1 on the $200/month plan to be around $2000 of API equivalent use.
> $200 of usage on a $20/mo plan is a 10X multiplier. […] So $200 * 10X = $2000/month. I don't know how you got to $8000.
FYI there are 4 weeks in a month…
> Recent analysis by industry analysts like Semi Accurate have measured Sol 6.1 on the $200/month plan to be around $2000 of API equivalent use.
Sol isn't an Anthropic product though so maybe OpenAI is less generous than the former, I can't tell.
I get a LOT of mileage out of my $20 plan. It's at least $500+ a week at API rates. I sometimes wonder if there's something wrong with their billing system.
On the $200* consumer plan, according to "npx ccusage", I'm using $11,000 per month at API rates. Possibly significantly more, because I was previously using the $100 plan so that only reflects about 1 week of the $200 plan.
I was spending thousands of dollars a day in funny money (free promotional tokens from VCs) before switching to a paid subscription for convenience.
You were spending zero dollars a day to buy tokens at prices that almost nobody actually pays? Woah cool story bro.
i don't understand why so many people on this site comment on issues they clearly do not know anything about
Obvious bullshit that's been disproven many times already.
The price is the price, everything else is bullshit.
> Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
You have confused necessary and sufficient.
> Also, Anthropic is showing that to operate at Anthropic level, you need $300,000 compute spend per employee
There are a bunch of open source harnesses that are competitive with CC and they aren't using anywhere near that many tokens to do it.
It's not just the LLM costs, to keep your agents busy you need to move them of your workstation onto a server/cloud. I currently have about 16 claude code agents working on different bugs, uses so much resources that my macbook pro m5 max becomes slow to respond. Using cloud session would free up my macbook for other work.
Anthropic recently handed out free cloud resources (it's still free, guess they want you to get used to the fact that cloud is not free). I went through $250 cloud budget in less than 2 days.
Feels almost as expensive as LLM calls.
I'm impressed you have a workflow that can keep 16 separate agents fed without the whole thing collapsing into nonsense. Even running full agent teams I can't think of ways to do that without being intentionally inefficient
I let one agent categorize all bugs, then have it decide which clusters of bugs should be addressed together, and which ones can be addressed in parallel.
Then all the other agents get to write more bugs.
You call your laptop "workstation"?
Technically it is... According the original vision for workstation, a computer with >1MB RAM, >1 megapixel of screen resolution, >1 MIPS and costing under 1 Mega cents (10K dollars).
However, so is my phone!
https://en.wikipedia.org/wiki/3M_computer
My laptop is shockingly competitive with my quite nice workstation.
(IMO the really impressive part is the thermals. My laptop can dissipate some 45W on a sustained basis, and my workstation takes quite a bit more than 45W to outperform my laptop.)
6 years ago i had a lenovo thinkpad p1, branded by lenovo as a portable workstation. Today i have macbook pro m5 max, much faster and more powerful than the lenovo, so why not call it a workstation?
just get a mac mini headless
Wouldn't that depend on how much hardware you have to run them on, too?
So every business should measure any client fungible resource that way? Should a bank put a revenue number to any capital held in excess of regulatory requirements, for example?
Excess capacity sitting idle is the cardinal sin in modern economy. I was merely pushing back at the insinuation that internal costs should be used to measure internal resource use. No. The costs should always be external market price because that's where you are making the correct trade-offs.
Bank reserves, Compute etc are trade-off decisions. Buffett vs Mayoshi Anthropic vs Google are being played out in real-time for us. Pretty soon you'll know which trade-off was better.
Zero excess capacity makes things quite unstable and estimating what revenue could potentially be had isn't necessarily easy or so reliable (and what about future revenues from using capacity internally vs revenued now?).
There is a huge difference between Anthropic and Google dog-fooding: Anthropic does not own or build data centers, they rent (as far as all the reporting on this I've read). Anthropic employees dog-fooding with Claude tools isn't "internal", its most likely being run on external AI data center compute, which they have to pay for, there's just no margin on it. It might even be run out of Google's data centers.
>> Anthropic does not own or build data centers, they rent
Thats a problem since they will always pay a pemium for compute. The data centers owners can compete on price if they choose to.
> So, every compute Deepmind and Googlers get is lost revenue opportunity and must be measured as such
Maybe, but it also reduces supply that can drive up prices
Is there any reason to believe that if this computer wasn’t used by an anthropic employee that it would be sold at retail price? That would imply that Anthropic’s computer is 100% utilized at all times, with no slack - something highly unlikely.
It depends on if Anthropic make more money on their internal spend on tokens than selling them externally. I strongly guess they do.
Not really- since they have an internal view of low and high usage times of the day, they can optimize differently.
Even within a single moment, if there is capacity for an internal request, just send it, if market traffic is currently at 100%, don't.
If it were opportunity cost, that would imply that Anthropic can't sell as many tokens as people want to buy. But as far as I can tell, there aren't any "token shortages," at least at API prices. Or is there a maximum per-customer limit that I'm not aware of?
By 'operate at Anthropic level', you mean terrible reliability and availability, and terrible efficiency right?
aka Cobblers children have no shoes. Anthropic profit less by selling it.
> No. This is opportunity cost.
This is NOT the definition of opportunity cost. Please do not throw in random terms you don't understand. Opportunity cost is the value of the next best alternative you give up whenever you make a choice. WTF does it have to do with mis-classifying an expense?
[dead]
Plenty of companies use retail/wholesale pricing internally to make sure a business unit doesn’t appear profitable just because it’s being subsidized by the one providing the resources. Plenty also have to have arm’s length pricing for legal reasons.
Anthropic isn’t booking it as GAAP revenue, they’re just using the number for marketing.
My understanding is most use a cost + figure and not retail. Do you know of any that use retail?
That's not really true. Measuring the equivalent retail cost of the usage is the only way to put it in a relatable perspective.
If Anthropic had revealed their internal cost, there comments would be complaining that they were hiding something and giving useless, unrelatable data.
The expenditure is most interesting when compared to retail rates, because that's the rate the rest of us have to compare against.
What's is the 'retail cost'? Because nobody (except total idiots) is actually paying their published API pricing.
Most "enterprise" companies are paying full API costs. The subscription deals are only available to consumers. If you want enterprise levels of control over how your staff are using the models you need to pay up.
Not even close. "enterprise" companies are negotiating contracts. Call it pre-provisioned capacity, wholesale pricing, etc. Same is true for AWS, Azure, Salesforce, or literally any SaaS provider ever. There is a reason literally every single pricing page for any SaaS provider has an "Enterprise? call us" column.
From what I've heard those enterprise negotiated discounts for Anthropic tend to be in the 15-30% range, which isn't anywhere near the deal you could get from using the $100/ or $200/ or $500/month subscriptions.
that still directly contradicts what you said above...
I said "most", not "all". I'll take that risk.
Do you have any data to back that up? Not really worried about which pronouns are used in this case
I do not have hard data that can support my claim that "most" Anthropic enterprise customers are paying full price. Just anecdotal evidence from stories like these:
https://blog.pragmaticengineer.com/the-pulse-token-spend-bre... - "I talked with companies spending $5M+ per year on Claude which have received no discounts. If Anthropic offers discounts, it will likely be at a much higher tier."
https://www.theinformation.com/articles/anthropic-openai-fig... - "Anthropic’s discounts on its models can be around 15% off listed prices, according to managers from the three software firms that buy AI from both companies."
Plus my own personal conversations with a small number of people at medium to large sized companies who said that Anthropic's sales teams were hard to get discounts from.
[dead]
The enterprise discounts are not as extreme as the personal subscriptions.
Unless you managed to talk them into that, in which case, share your secrets!
Really? How much of anthropic's revenue is coming from API usage at public/retail rates (without discounts or credits)?
That information isn't public, but the closest we can get is from the S-1 telling us that for the nine months ending September 30, 2026, consumption-based (pay-per-token) revenue was $26.3bn, subscriptions were $6.9bn.
Page 140 of https://www.sec.gov/Archives/edgar/data/2130357/000162828026...
If that information is not public (of course it's not lol), I don't see why you would follow that up with stats that don't actually answer the question and imply they somehow get us closer to the answer...? do you own anthropic stock/options?
No, I deliberately don't own stock in the companies I write about (or any single companies at all, aside from a Vanguard index fund.)
I've learned enough that I am confident that enterprise companies pay a lot more for their tokens than $100/month subscribers, and that negotiated enterprise discounts don't change that fact in a meaningful way.
I don't disagree that they pay more, but how much more? The value/usage of a $100 sub is also fairly opaque and perpetually decreasing...
[dead]
Big companies actually are paying published API rates, Anthropic and OpenAI models are desirable enough that there is very little wiggle room to negotiate. It's their way or the high way. Same with things like "safety" filters and (for Fable) mandatory data retention.
Most of their revenue comes from customers pay API pricing.
Those individual subscription plans you get as a consumer are not what their big enterprise customers are using.
Those plans are cheap so you, as an individual, can get familiar with the product and convince your company to get it at work. Your company pays the API prices to get enterprise features.
i am sure there are companies that push individual employees to get the cheaper individual subscriptions
In the US, the IRS doesn't see it that way, at least when it comes to fringe benefits as a part of employee compensation. https://www.irs.gov/businesses/small-businesses-self-employe... says, "In general, the amount the employer must include is the amount by which the fair market value of the benefits is more than the sum of what the employee paid for it plus any amount that the law excludes."
As an example, as I understand it, if a hotel chain allows an employee to stay in one of their hotels for free when they go on vacation, the tax value of that benefit is what they would have charged someone for that hotel room, not how much it increased the hotel's costs to have one more hotel room full — which is generally almost nothing.
Of course, the IRS does agree with you when it comes to expenditures that are for the employer's benefit, rather than the employee's. If it didn't, almost all companies would have negative profits from the IRS's point of view.
Not an apt comparison. You are conflating the opex of the company with taxable benefits.
If what you said was how taxation worked, a hand tool manufactuting company giving hammers they make for $1 and retailing for $2 to workers building a shed on their property would be increasing the workers' income tax bill by $2. Of course, in reality the workers get no income tax bill on the hammer given to them for work and the company can get a $1 tax writeoff from the opex of making the hammer.
> increasing the workers' income tax bill.
They are not giving anything to employees for their personal use to use as they please, it is not part of their benefits whatsoever. It is work tooling, paid for by work. This is akin to Microsoft paying for Github or O365 for their staff. they can hardly put it in benefits, it has nothing to do with employee compensation or their taxable benefits.
OP is just using the benefits analogy as a comparison how IRS typically accounts for similar things when it is a benefit.
Employees staying at hotel for their personal vacation is a taxable benefit. Employees staying at a hotel as part of their job is a perk, and not considered a taxable benefit - A cruise ship can hardly say cost of cruise is now part of employee benefits (for tax it is not part of your comp) . They can of course market it as a "perk" to attract talent - same as the armed forces using cool extreme sport videos as an talent attraction tool. You can use in marketing yes, but it not part of your compensation.
Anthropic and other Frontier labs /Big Tech can definitely pitch token allowances as a selling point part of recruitment, but it cannot be counted as a taxable benefit - unless the employee can use it as they please outside work for any personal reason.
IRS rules should never be used as an example of how accounting "should" work. They're full of politics and the implications of the rules on tax avoidance.
Suppose a company could transfer an asset to an employee and have its value set as the employer's cost. A company could make a bunch of volatile investments and see which ones pay off. Then the ones that go south get sold back into the market as a tax loss and the ones that do well get transferred as compensation to employees at the purchase price instead of the current value, reducing the tax liability of both the employer (no taxable gain) and the employee (less taxable income). What would happen if any employer could do that with any employee and any asset?
Conversely, there is a separate carve out for "employee discounts", which is essentially the employer offering its employees its own products for something in the neighborhood of its own cost. Why? It's not in any way a principled accounting distinction, it's because it's harder to use for tax avoidance at scale, because then a soap company can give its employees a discount on soap, but people only need a finite amount of soap and then the soap company can't use it for cars and real estate and stocks.
Suppose a company could transfer an asset to an employee and have its value set as the employer's cost.
That's not how it works...The value would be the FMV at the time of the transfer to the employee. The cost basis would be the employer's cost. The income to the employee would be the net difference.
What would happen if any employer could do that with any employee and any asset?
They can't, because tax laws have addressed this situation for almost 100 years...
there is a separate carve out for "employee discounts", which is essentially the employer offering its employees its own products for something in the neighborhood of its own cost.
That carveout only applies to products being offered for the employee's own consumption unrelated to the needs of the business. Internal use of a company's own products for the needs of the business does not result in additional GAAP or tax deductions (meaning, if that product cost X to make/provide but they charge external customers Y, the only deduction is X, not Y).
> That's not how it works
I know that's not how it works, I'm explaining why they do it in a particular way because of what would happen if they did it the other way.
Your AI token bill isn't a "benefit", it's just an expense, just like all the other software subscriptions your work pays for.
It might be difficult to value because datacenters have high fixed costs and upfront investment and low marginal costs. If I spend 1B building a datacenter and 10M/month to run it (assume this varies negligibly with usage) and on average 25% of it goes to internal use and 70% to customer inference and 5% idle, should I charge 25% of it to internal use even though the marginal cost is lower because internal workflows are predictable and can be scheduled during off peak hours? With that being said I agree it's wrong to value it at the retail price (unless they are supply constrained) and this is probably for marketing purposes.
> Only a company like Anthropic
No. I worked at an airline where the company typically charged itself higher than retail for tickets. At least, it was the highest fare type.
I worked at a media company - we had to pay top rates in order to buy our own ads for cross selling products.
Finally, I work at a SaaS company who heavily uses our own product. While we don't pay for it - just our cloud hosting bill - it's definitely a discussion point about how we should be a bit more mindful about usage.
I’ve worked at and with large orgs - it’s absolutely standard practice for BUs to bill other BUs at “retail” or at the very least a commercial rate.
Absolutely not. Any company that moves from API to self hosting (Hello Revolut on the frontage) will publicize the savings, e.g, the actual cost vs API pricing. You might even do that if you start out self hosting.
Microsoft basically does the same with internal Azure services? When I was there our budget/usage was based on and consumed at the consumer price for GPUs and other compute. What happened behind the scenes really I have no knowledge.
Not true. Azure has a very different internal pricing. You can often get a glimpse of it if you have a.. particularly enterprising TAM. What happens sometimes (Like now when compute/GPUs are in high demand) is finance department would calculate the opportunity cost on allocating those resources internally vs selling them externally. The calculation isn't even always straight forward. What is the opportunity cost of using compute for internal CI or company operations vs selling them externally? It's not a straightforward equation and you would have to adjust the scale often depending on how clogged your internal processes get.
The extremely discounted internal cost promote waste, so there is always a room for improving efficiency, but that too has a cost.
That makes sense! As an IC all I saw was the public number.
It's exactly the other way around. Anthropic is running these models at a loss if they charge consumer price. The real cost is much higher than they claim here.
In a lot of companies, you would get charges back on usage/license/cost and this can get even funnier in NGOs.
In a similar way, internal employees would have also a "standard daily rate" like an external employee.
Anthropic might well be actually spending more than their retail prices to host the API so not sure going to actual cost would improve these numbers.
They did lose an insane amount of money last quarter versus revenue.
It doesnt work like that internally. It would be different BUs purchasing from other BUs. Ir probably helps with taxes.
I think it's usually cost +, but yeah, I don't know of another company offhand that sells the services at RETAIL.
> You blow $50k a year on servers running your LLMs? It cost you $50k.
Plus electricity
Seems entirely in line with Hollywood-style practices.
> cost they charge customers
price they charge customers
but they're using other people's data centers, not their own
Do they actually book the revenue against themselves?
This is the world of spend more, get valued more; obviously the investment will be used to spend more and get more, so it must be good. This is investor-think. If you don't have spend, why do you need investment?
(Or it could be bad. Very very bad)
Actually I think it's very informational. For every other company, working on the cutting edge of technology, paying for intelligence at retail prices, means you are now paying more for AI than you are employees. That $ could have been spent on 2x the employees, however that's not as useful as 50% AI / 50% humans, and indicative of where we are headed.
>means you are now paying more for AI than you are employees.
Yea, isn't the entire point of RSI to replace your employees altogether? So you'd spend an infinite amount more on AI than employees in that theoretical.
Like, if you were in an energy generation business spending way more on fuel and capital expenditures is way better than having zillions of employees earning checks , it would be a sign of efficiency.
Anthropic's use of AI in no way translates or represents any other companies' use of AI, so your apparent extrapolation makes no sense.
Anthropic is full of the smartest talent and a bellwether for how the rest of the service industry will work. Just as cutting edge companies of the past, like 2005-2020 Google's use of computers and software were indicative of the software every other service company was going to be using, lest they turn into an old fashioned uncompetitive dinosaur.
Wikipedia tells me that "service industry" is a term that includes transportation companies and law firms.
So you are apparently now claiming my local train company or law firm will be using as many tokens per employee as an A.I. lab.
Yes service industry is quite broad, is white collar work more clear in general? Yes, it is quite obviously apparent that law firms are and will be using lots of token-based intelligence. They won't be using the same tokens at the same time as an A.I. lab. However, the best law firms of 2029 will be using approximately the same tokens per employee as an A.I. lab of 2026, that much is clear.
> . However, the best law firms of 2029 will be using approximately the same tokens per employee as an A.I. lab of 2026, that much is clear.
This is as if you learned that Anthropic scanned 1000 books per employee this year to train their models and your conclusion was that the best law firm will be scanning 1000 books per employee by 2029.
You still pay input tokens for Jev. "Should I crash this train, YES/NO?"
Cool story bud
[dead]
The industry-wide curve looks similar. OpenAI's API went from about 6B tokens/min (Oct 2025) to 15B/min (Mar 2026), and OpenRouter from 113T tokens/week (Aug 25) to about 168T (Oct 5). I've been charting these against the ~100T words humans produce per day at aiwords.si
It's kind of expected, no? If I was an Anthropic employee expected to dog-food the product and be on the frontier of usage, I'd probably be using Mythos all the time at the highest thinking level. And if I am not being metered, I'd probably do more stuff that I know are likely low yield since I am not worried about a quota. This would probably take my "API-equivalent" usage beyond the average salary in far less than a month.
I wonder how much dog shit has crept into their codebase by this point. Another company might dogfood their product but Anthropic is so heavily marketing oriented, I doubt this is driven by the desire to improve the product as much as it is by the desire to pad out usage numbers and create shareable stories. No matter the consequences. I bet $1 it sucks to be a good engineer in their software effort.
I'm personally really not impressed by their software quality, at least we know from their experience that infinite tokens at the best model available still doesn't replace good engineering.
It doesn't replace MBA-brained bad roadmapping. Good engineering is determined more by the processes you have in place for catching something that breaks bad. The system produces good code - singular 'good' developers are great but if they're not socializing what makes them good then it's just competency. Great engineers teach and mentor and advocate for capacity dedicated to testing and health-checks.
Given the speed at which Anthropic regularly pushes updates (eg their ridiculous app versioning that's in the thousands for v4 now) I doubt they're doing anything that looks like quality processes, even AI-integrated ones other than basic code review skills.
I imagine it's as much of a mess internally as it is with any other company whose execs are concerned more with bottom-line valuation rather than maintainability, extensibility, and sustainability. Just lots of spaghetti-at-the-wall hoping that something sticks long enough to drive up the IPO.
I find Claude to be exceptionally well-designed software that almost never has notable issues.
On the flip side, software I used that was majority built by 'real human engineers' coding by hand is utter garbage.
As a recent example, Facebook Ad Manager is extremely slow, plagued with UX issues, blatant state management problems (double scrollbars, iframes, random rerendering).
Other examples of 'real engineers' building software include SAP, Oracle ERPs, most banking software, most government websites. They're trash.
Human engineering is the reason your personal information gets leaked every other week. Human engineering is the reason there's a critical vulnerability in core software every day at this point.
The average software engineer cannot build good and secure software. Just like the average car driver is a safety disaster on the roads.
Claude is an npm package, a terminal app. A 1/2 million lines that probably translates to around 50-100k lines of original thinking. A pair of industrious devs (or one inspired maniac) could write about 100k in the first year of a fresh project.
That is one way to use it, but there's also the Claude desktop app that is vastly more feature rich.
>>If I was an Anthropic employee expected to dog-food the product…
>I wonder how much dog shit has crept into their codebase by this point.
You'd think they'd know not to shit where they eat… ?
A while ago the Claude code source code was leaked. It was very revealing.
From what I remember they had a large regex to match “bad words”
Rather than using their AI to simply scan messages to flag.
Computation though is still an economic problem of applying compute resources to information processing problems. Regexes that work are much more efficient than a massive LLM inference.
> employee expected to dog-food
I guess kinda put you in a bubble a little, I wonder how the experience of Anthropic eng who left the company would look like, when such eng would go and spend few $K a day in non AI lab, would that be encourage/allowed/taken as example?
> If I was an Anthropic employee expected to dog-food the product and be on the frontier of usage, I'd probably be using Mythos all the time at the highest thinking level.
Mythos at the highest thinking level is not the only product from Anthropic though.
True. I was reacting to the article mentioning that April was when the usage exceeded salaries. Mythos was released in restricted preview through Glasswing in early April, and I presume was available internally for at least a few weeks before that, which helped tip the numbers.
> I'd probably be using Mythos all the time at the highest thinking level.
Especially when I need to plan my next hike up Mount Shasta.
My fist reaction to this post was "why so little?". I very much expected the number to be higher.
I'm not really sure I understand the excitement around this. It seems like the implication is that if retail price > employee salary, you are replacing employees with AI spend? Yet, anthropic's head count has nearly doubled since last year - so, in a company where nearly half of that head count is engineers, but claude is "writing 90% of the code," what is everyone doing then? You would think, if the hype feeling around this held true, that the output would be going up and costs remaining stable. Costs rising as you scale is not a new thing but that's not really what AI promises. It sounds to me like the retail AI spend per employee is very high. What is the appeal of that?
> what is everyone doing then?
Making sure Claude doesn't kill everyone on Earth, if they are to be believed.
Imagine an employee is 1x productive without AI.
Now spend 0.5x their salary on AI, and they become 5x productive compared to no AI.
Now spend 1.1x their salary on AI, and they become 20x productive compared to no AI.
It's still obviously worth it to spend the 1.1x. But it is truly remarkable that the biggest cost of hiring a knowledge worker is no longer their salary, but their AI usage.
There is no implication here that employees are being replaced by AI spend, as long as all of the productivity is generating profit. The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself, so you still need the employees.
> Now spend 0.5x their salary on AI, and they become 5x productive compared to no AI. > > Now spend 1.1x their salary on AI, and they become 20x productive compared to no AI.
This is what I am getting at. If this were true - and I've yet to see evidence it is, despite some anecdotal evidence (organizationally, IME, it's about 2-3x for power users, and probably a net loss of productivity in certain areas) - why would their headcount be growing so much still YoY, pretty much doubling every year. If the AI is 10-20xing why such a big jump in headcount? Yes, the numbers are shiny, the growth is apparent, I would just expect with this AI narrative they'd be "doing more with less" as this thing improves every year.
> The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself, so you still need the employees.
This is the story AI sells though, and Anthropic has participated in it heavily. Yet, their interview process requires hard skills and doesn't let you use AI, so clearly, what you're saying is correct, it just doesn't really align in my head with the whole knowledge worker being replaced narrative.
> I would just expect with this AI narrative they'd be "doing more with less" as this thing improves every year.
That wouldn't really be an expected outcome for a company like Anthropic. They need to do the maximum they can with the funding they have, or risk losing out to competitors who don't take their foot off the gas.
Jevons paradox strikes again. When work becomes more productive often the business finds so much more to do that they end up using even more resources. Why do more with less when you can do even more with more?
You could always do less with more: https://ctolunchnyc.substack.com/p/doing-less-with-more
I mean, just what I did this morning with Claude would have taken me 2 weeks on my own. So 20x is pretty reasonable for certain types of engineering.
But it's also true that when I'm working on 2 things at once, the bottleneck is me. I'm spending all my time giving Claude direction that it can't give itself.
So you basically have to separate out the tasks where AI can produce a 20x productivity boost, and the stuff where it can't help at all.
There's still so much stuff where it can't help at all, you need real human employees for that. That's why Anthropic is continuing to hire at breakneck pace.
And that's what I'm saying -- knowledge workers aren't being replaced. They're just accomplishing tons more work, and companies aren't running out of work to do.
Sorry, I know this is the story, but am I expected to believe that Anthropic would need 20,000+ engineers in their org without their use of AI? I do not believe that is close to true.
No, one thing that would happen is that many projects would never be undertaken - basically, there's latent demand for software that would simply go unmet, because of resource constraints.
I was recently involved in a case like this - we had a critical part of our system based on a dependency that's no longer maintained. Previously, we would have just carried on for a year or more, living with the limitations until it was much closer to being entirely untenable, because rewriting the code in question would have been a multi-month project involving multiple developers, QA, and management time. But now, the code in question was migrated in a few days using AI, along with a suite of new tests in addition to our existing ones, that gave us a lot of confidence in the results. The productivity improvement here is at least on the order of 25x, possibly higher.
But most developers aren't sitting at their desks operating at a 25x pace all day, every day, so average productivity improvement is going to be lower. And of course, some use of AI will be suboptimal or problematic: waste due to e.g. duplicate development which happens when development is so easy, or costs for cleanup of mistakes.
We also now have a slew of AI-generated internal tools, most of which would never have been developed before. Many of them are very useful, the kind of thing that's on your eternal wishlist but never gets done. Some are kinda useless, so represent wasted resources. It's difficult to measure the productivity benefit of all this, because you'd have to compare it to the non-existent company without those tools.
Ultimately, measuring all this meaningfully is not really that feasible, except perhaps over the long term. Managers operate on vibes with this kind of thing because in many cases, that's all you can really do.
They are 20x productive on the old KPI.
And the product now iterate 2 times faster with 10 times more bugs and customers are pissed.
Looking at you Apple: https://aperocky.com/blog/post.html?slug=apple-watch-sucks
Yes, one can quite easily make quite fantastic numbers fall out of simply applying arithmetic and not dealing with messy propositions like grounding those numbers in any kind of real world outcome.
This is the guy you want to be asking about the emperor's taste in fashion.
> The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself
So we're back to square one, and your productivity increase has a ceiling that's much lower than 5x, let alone 20x. There's a limit on how many things one can keep track of, how many code changes one can reason about.
I'm not really following. My point here is that this is why the ceiling is around 20x, as opposed to being e.g. 200x or 2,000x.
That doesn't mean it's "much lower than 5x."
Now spend 1.2x their salary on AI and they become 99999x productive. So why not just go further if we're making up numbers?
> In March, I wrote about a new trend: Tech companies are starting to treat AI tokens (the main way to measure usage) as a fourth component of employee compensation, alongside salary, bonuses, and stock awards.
There's a lot to unpack about this statement.
Until employees can use those AI tokens for personal projects, it's not compensation. Rather, it's just a cost of doing business, like laptops, office space, or testing servers.
Do we think they're not using them for personal projects? It's called "testing." ;)
Does the code that their corp AI generated belong to them ?
I'd say it's a "don't ask, don't tell" situation. Perhaps they monitor the sessions though and might wonder why someone is working reverse engineering a retro video game.
> There's a lot to unpack about this statement.
Including the fact that it's probably not true.
If something is compensation, it can often be negotiated and traded off against other compensation components. How many people have successfully negotiated AI tokens away in favour of salary/bonus/stock? I'm going to guess "approximately zero".
Yea they used the wrong word and it felt like rage bait. AI tokens you use for your job given by your employer is not compensation. Hell money isn't even compensation in many cases, it all depends on the context. The author needs use AI to help them rework that sentence or open a dictionary
Is that really true? At least at large companies, the components are not usually fungible in my experience. At startups or smaller companies I'm sure it's a lot more flexible.
Even if the request gets rejected, it wouldn't sound completely off to anyone if you tried to negotiate more salary for less shares.
I think almost all companies would laugh at you if you tried to reduce the number of tokens you can use for more salary. The tokens are stuff you can use to be productive at work, I expect that companies that take tokens seriously enough to mention them before you start wouldn't like that at all.
I think the author used the wrong word.
I haven't seen anyone treating AI tokens as employee compensation. If anyone can show me an article where employees were paying taxes on their token use at work, please share it.
Companies have been treating AI tokens as part of the cost of having employees, like the cost of their office or travel expenses.
> I think the author used the wrong word.
True, though that phrase understates the author's problem a bit, as it is surrounded with more wrong-words, indicating it's a wrong-idea rather than a slip of the tongue.
As you note, it's closer to office/travel expenses. Employee computers, licenses for software on those computers, licenses for SaaS stuff, paying for bandwidth, etc.
It's certainly a cost of having an employee, along with salary and benefits. But employee costs also include things like laptops, electricity, internet, management, support staff, and possibly office space. Tokens are much more like the latter than the former.
Many companies used to brag about about how nice the laptop or office was as a way of convincing people to work there. Often in the same breath as compensation.
But compensation is only one component of cost.
So I think this guy is confused about what it costs to have an employee and what the employee gets out of it. Possibly intentionally.
From the point of view of the pay department, maybe. From the point of view of the employee, absolutely not.
It is like counting tolls as a component of "employee compensation" for a driver. An employer may decide to have a driver use toll roads, which may result in faster deliveries and more efficient use of his time. From the employer point of view, he is paying more to get more work done, from the driver point of view, he is paid the same for the same amount for time driving.
I can see where there might be a truth to it. My company offers a fairly generous monthly token budget. If I were evaluating an offer from another company, but they offered a lower monthly budget, it would definitely influence my decision, all else being equal. It's not compensation, but it does affect my day-to-day experience at work.
"We have great tools and you'll get to work with cutting edge technology" is a selling point!
In case anyone wants to see the original source, it's page 187 and 188 of the S-1: https://www.sec.gov/Archives/edgar/data/2130357/000162828026...
Under the heading "Claude is Accelerating Our Ability to Close the Loop":
> The expansion of our own use of Claude reflects this leverage directly. The value of our own internal Claude usage at list prices has grown by over 250 times from January 2026 to September 2026, and in April 2026, this monthly value first exceeded our monthly personnel expense, excluding stock-based compensation. Because each version of Claude is used to help build the next and because Claude makes our teams more productive across the company, we believe the compute we devote to internal usage compounds these gains, continuously accelerating our pace of development and further strengthening our leadership position.
Dumb question, what most engs at Anthropic work on? It doesn't seem like release some large new products all the time while they have quite a few of engs. What are the major branches of work besides LLM research? I guess there is a lot of infra stuff to support the scale of operation, just supporting API and TUI shouldn't be that demanding I guess.
- Cross-platform desktop apps - Cross-platform mobile apps - Claude Code TUI - Product-specific verticals like Claude Design - Corporate marketing website - APIs - Writing technical documentation for APIs - Enterprise support - And what must be a massive infra team to run training and inference at scale in a heterogeneous cloud environment
And I’m sure more than that. But there’s obviously plenty to work on, and they’re not at the point they can just Claude their way through it.
I'm kinda curious about this too. But then I'm always amazed at the enormous workforces these tech companies have, and I'm sure they cut them as thin as they can. Like adobe has 30,000 employees & some blogger estimated that 1/3rd of their workers are actually working on their software. I use their software sometimes and I really haven't noticed many changes to it over the years, except recently with the ai stuff. Like from 2015-2020 I can't think of any major things they added to adobe illustrator. I guess making and maintaining software is hard.
I fully agree most big tech companies are pretty bloated. But FWIW a lot of my time at SaaS companies, especially in the B2B space, was spend working on features for larger customers that most others would never use or even see. Like, yeah, product X isn't that complex, but $BigCorp wants to use it in their private cloud, or needs it to be single tenant, or needs these audits done, or want their $Gov client to use it. Now you have a ton of work for zero visible changes.
Good point. I'm sure a lot of the work adobe was doing between 2015-2020 involved moving to the cloud and stuff like that which I just never interacted with, and cloud stuff probably has a lot of B2B work. Plus during that time the value of their stock grew like 500% with an ATH at 700%!! They were kicking ass! Probably, to justify their growth they had all kinds of projects they were working on, some of which never saw the light of day, others which grew new parts of the company like Adobe Dimension.
Even at companies that appear super bloated I don't think people are idle, its just that mantaining stuff is a lot of work, projects get cancelled, and, like you said they do a lot of specialty work for big clients.
Adobe has a shit ton of product lines, people think of it as just the core Photoshop, Illustrator, etc., but just go look at how much random stuff there is: https://en.wikipedia.org/wiki/Adobe_Inc.#Products
That's in addition to running the big SaaS offering of the creative cloud and all of the other SaaS's (behance, adobe express, adobe stock, etc.).
All of those include software engineers, SRE, support staff, PMs, etc.. Then there's the massive amount of sales staff, go to market, etc.
There's also a lot of R&D and domain experts, stuff like Adobe Firefly? That was spun up in like a year apparently, probably continuing existing in-house research.
Claude for Legal
Claude for Teachers
Claude for Government
Claude for Financial Advisors
Claude for Clowns
Claude for OpenAI Employees
Claude for People Called Geoff Who Were Born In June
It's a long list.
I'd guess they have a lot of internally facing tooling for monitoring, one-shot eval to catch people prompting bad things, etc. That stuff can probably be yolo'd and it just annoys the internal teams using it.
After anthropic lost $1.5 bill for pirating books they decided to pull a google and manually scan every book ever written. The number of people needed to gather the data they need is probably massive.
Everyone is saying it is expected, and I agree, but it’s also clear that employees of AI companies spending unlimited zero-cost tokens in their day-to-day work is resulting in products that aren’t usable by the vast majority of regular workers who don’t have this luxury. Dogfooding your product is good but you should also use it with the same constraints that you place on your users, otherwise what signals are you really getting?
They're dogfooding the way Xerox PARC was dogfooding in the 01970s.
By giving hundreds of workers each a powerful desk-sized Alto minicomputer for their own needs, with a high-end graphical display, they gained insight into what desktop personal computing would look like 15 years later, enough insight to shape it profoundly. It's been 55 years, and although multitouch screens are proliferating, I'm still typing this on a keyboard hooked up to a framebuffer-driven display with a mouse hooked up to it, displaying hypertext, festooned with windows, icons, and a mouse that moves an onscreen pointer at the same 135° angle they settled on 55 years ago, positioning a vertical line to indicate the insertion point, with scrollbars. The display has twice as many pixels, the pixels are color, the mouse pointer is white rather than black, the screen is a flat LCD, the scrollbars are on the right, and the cursor blinks, but it's very recognizably descended from https://smalltalkzoo.computerhistory.org/HOPL-St78.html rather than, for example, ITS or GENESYS.
Only people with the luxury of unlimited zero-cost CPU cycles had a chance of seeing what the future looked like, and that allowed them to find out what worked and what didn't work in that environment. Nowadays, the regular workers who don't have the luxury of unlimited zero-cost tokens are manually reviewing voluminous AI code, because, even though that's a lot of slow work, they're still much faster programming that way than they were without it. Anthropic, instead, is doing insane things like eliciting a billion lines of Lean proofs to show that existing open-source libraries can't fail their OSS-FUZZ tests — and fixing all the bugs that effort turned up, in the process.
It’s true employees at the big AI firms are living in the future, but since this disclosure is in marketing materials to sell their stock to investors, it’s important to note that almost all the money in personal computing was made in mobile which didn’t use any of these ideas, and in desktop computing the money was made by Microsoft who treated the GUI as ancillary and competed on distribution.
Xerox made quite a bit of money selling laser printers for desktop publishing, too, a lot more than they ever spent on PARC.
I don't think it's accurate that "almost all the money in personal computing was made in mobile". Quite a bit of money was made in personal computing before the iPhone, and quite a bit of money has been made in non-mobile computing since the iPhone.
I also don't think it's accurate that mobile "didn't use any of these ideas". Android's UI is not quite as close to Smalltalk-78 as what I'm using right now, but you still have pointers, scrollbars, text selections, proportional fonts, vertical-line cursors, buttons, menus, object-oriented programming, sometimes a bytecode interpreter, and even windows. You just don't have as much freedom with where you position the windows. Android's UI is still a lot closer to Smalltalk-78 than it is to, for example, acme, GENESYS, the ITS HACTRN, FORTH, or PDP-11 Unix.
They aren’t usable? I’ve been using them a bunch and have been having good results. Dogfooding is always good, but I don’t think there’s some kind of major problem here.
So...given how much they are loosing and the looming bankruptcy risk, the lesson is: don't do this?
Are you kidding? Did you not see Anthropic's S1 filing? Never has so much value (in dollars) been created by a company in such a short time in the history of humanity. I'm not sure how your takeaway from that is "don't do this".
There's unfortunately a lot of emotions and cope around this so you will be downvoted. There was a recent viral news article implying that Anthropic had lost 80B in a single quarter, this was really because of increased valuations. But a lot still believe this is the case and Anthropic will continue to have this 80B loss..
The reality is that they are profitable in the normal sense
I would take this with a grain of salt.
Recently I read about a company (nameless for reasons that will become clear) that was giving a small team unlimited tokens to work on a very ambitious project. (IE, they'll probably spend a few million on tokens.)
Here's the thing: I once worked with the (nameless) exec who was named and interviewed in the article. (And I have a significant amount of respect for this exec.) The problem is, the exec (when I worked with them) significantly misinterpreted the mobile trend and nearly killed the company jumping on the mobile bandwagon. (We had a desktop product that only needed a simple companion mobile app. This was painfully obvious at the time.) This exec is blindly jumping on another bandwagon again, wasting money that might be better spent elsewhere.
Why do I say this? I question if a lot of big AI spenders are really getting value from what they're doing with AI, or is it just for show?
Friction is good. I've been in this for fifteen years, and I can't name a company whose product would have been better if they somehow had a magic wand to instantly implement even half of those 5 years, 600 tickets deep backlogs.
> Friction is good.
The nameless exec said that to me the last time we met, roughly a week before they were pushed out; precisely for that belief. (And at first I suspected you were the exec.)
And yes, the product that I worked on did have a ~600 ticket deep backlog, and plowing through it wouldn't have made the product better. At the time, the product was in a steady growth phase, where addressing customer pain points was just as important as building features.
Where the exec goofed (by jumping on the mobile bandwagon) was misunderstanding disruption, market needs, and how mobile would effect us. Ultimately, we were a "bolt-on" accessory, and the incumbents in our industry slowly pushed us out until our niche, (which had nothing to do with mobile) was our core differentiating feature.
Basically, the issue was allocation of resources. Too much were put into mobile, not enough into building features where we could compete against the products we were "bolting on to," and not enough into addressing customer pain points. The "friction is good" approach also severely hurt us, because too much time was wasted haggling instead of doing. (I specifically told the exec, the last time we met, it felt like he was running the air conditioner and heater at the same time.)
This exec is making a similar mistake by basically giving a small team unlimited tokens, instead of understanding how other players in the industry tackle these same problems with AI. It's not that their overall idea is bad, it's that they are probably starving the rest of the business on an unproven experiment / poorly chosen risk.
I’m not sure what this “stat” tells us other than Anthropic is dogfooding its own tech. Which is kinda expected.
The real question is what could they have done with open weight models on the same compute they’re renting. If the answer is “basically the same” then this flips the entire conversion back to what is Anthropic’s moat here.
I mean, they probably could not have done most of it with open models. Again, a lot of this isn't going to standard software development, at least I would assume but instead to SOTA model development and RSI.
Are a majority of Anthropic developers working on SOTA models and RSI? I would have guessed not.
There are many internal developers at $dayjob (where core products are not based on AI) that spend over a thousand dollars every day on average, which is likely higher than their salary.
I doubt they are bringing that much value to the company for whatever work is done with that much token.
that's crazy: in my experiments (and experience) i can get ~ 90% the quality of the most expensive settings on frontier models for far far less than 90% of the cost by changing settings down.
all of that to hit two nines of uptime
Oh is it finally up to 89.9%?
"ohhh, la-di-dah, this guy's got tens in front of his two nines instead of just 9.9" - github, probably.
Do the same math on AWS cloud use at AWS and you'll probably see proportionate massive costs.
This is the nature of managed anything. AI efficiency will improve but essentially this is what automation does.
"Tech companies are starting to treat AI tokens (the main way to measure usage) as a fourth component of employee compensation"
Why? My tokens are spent on improving or shipping features for the company. This is not comp. It should be added as a component of expenses, like a laptop, office, chair, etc.
Note, this doesn't include options or stock benefits. That's a huge part of actual compensation for engineers at this level even if it's on a different balance sheet.
Does anyone pay retail API pricing to Anthropic at scale? The last time I checked, I was burning > "$10K" per week with my $200/month subscription.
I know there's a point where companies get big enough, and Anthropic forces them over to the API, but presumably such companies negotiate heavy discounts instead of eating the 200x overnight price hike.
Unless you are a Broadcom and have leverage, yes. You may get 10-30% off but it isn't like a serious enterprise discount.
- @robobun, plz rewrite in Rust.
- @fable, plz s0lv3 Riemann. You are doing great.
- @claude. rewrite yourself for the 10000th time. Use as many tokens as you like.
At some point you're basically hiring cheerleaders.
So the ai tool cost as much as an engineer. Note that anthropic pays a much higher salary than the average tech company. So how are they selling these tools to companies? I understand time to market is the advantage, but why should the non-software companies adopt this?
> excluding stock awards
So it exceeds employee pay if you discount the vast majority of employee pay.
The title is as stupid as Dario.
Regardless of one's opinions about future AI usage, I'm not sure this metric says anything at all.
Would Anthropic themselves be paying this much if it was actually billed? It is not obvious.
Is Anthropic a good sample for AI use? Surely not, since they have special needs (dogfooding) and a vested interest in proving that AI usage is the future.
Can we compare their results pre and post AI usage? Close to definitionally impossible.
Are their employees good examples of AI-assisted engineers? Not really, as (one would hope) they're more well versed on the topic than an average eng.
This is like how soldiers use tools worth orders of magnitude more expensive than their own pay.
It’s not weird once you think about it. Money doesn’t create an operator. But operators can be cheaper than tools and often are.
If 90%+ of the code is written by AI, why do they need so many engineers?
Anthropic isn’t capable of not using AI anymore, so there is no off switch. Their AI could turn evil and they’d still have to use it. They are compromised by AI; it controls them now.
So, new development in AI is contributed by AI more than by humans. Isn't this the exact definition of singularity?
For every new hire, you have to ask yourself: "Will this employee be better than another $300k in tokens?"
I have thought of this and I think if you have value you can gain by adding more directing power (a person) then yes. Get another person to unlock that token spend on creating value. Otherwise, don’t hire that person as existing resources can direct that token spend. If you have no where to grow, adding another person will not benefit from more token spend.
How many tokens per employee usage monthly?
This is not a good measure. You have to measure productivity. If the AI is better, it DESERVES to be paid more.
What does this translate to in terms of number of 24/7 coding agents per engineering staff member?
as computing power grows and LLMs optimize their cost, that number is going to come down. I think it's temporary, and I'm optimistic about the future
Could it be because spending it internally means they can’t receive retail prices for the tokens?
Is it because it's "unlimited" ? An analog would be lifestyle inflation.
Don’t get high on your own supply
"Drug dealer smokes as much as he makes in a year." (news at eleven)
I would have expected like 10% human cost and 90% ai for any company that understands how to use it. And with their level of access, insight, and internal pricing, I would have expected the ratio to be more like 2:100.
Actually I don't believe their statement is true, they must use it much much more.
What is Anthropic actually building with all of that AI usage. Their only public facing products are Claude Code and the models themselves, are they just churning their codebase? Are there big internal projects?
They probably use their models to help build the training pipelines as openAI does.
They giving stock options to the AI now?
Or has it broken out of its jail and taken over the board? Dario seems like a bit of a windbag anyways, AI can probably automate most of what he does these days.
"Well, you know, you know, you know, I—I—I—I just—I just do what the machines tell me to now. Th-th-they're, they are smarter than us now, a-a-and it makes more fiduciary sense to let them have control."
Is their headcount increasing or decreasing?
They should use Deepseek - much cheaper.
this is meaningless trivia as paying retail prices to yourself is nonsense.
Now imagine if they hired from other geographies.
I'm sure in several US-based companies, non-US workers AI spending is already beyond their own cash comp.
> Anthropic's numbers show why. More than 90% of the code added to its software in September was written by Claude. Meanwhile, revenue has grown vastly faster than headcount.
Correlation does not equal causation?
ummm yes, they have to dogfood it... this is that lucky area of the venn diagram where you have to do it, measure it an act on it... because improves the bottom end and the product.
[flagged]
[flagged]
everyone who said "why aren't AI companies replacing their employees with AI?" should be rethinking their position
To what?
All this news says is Company A uses internally the resource they sell externally at ($X) value. It's not saying how much it cost the company to generate said consumed resource
to spell it out: the implicit claim is "ai companies think ai is a scam, as evidenced by them using humans instead of ai", and I believe the huge internal ai use is evidence to the contrary
AI, the technology, can be useful... And... AI the investment case can be a scam.
These are not mutually exclusive points, and I find it a common issue to get both sides to understand.
Anthropic/OpenAI should start doing mass layoffs if they want to convince other companies that this is really the human-labor killer.
they're using ai instead of employing more people, how is that not exactly the same minus a completely morale-destroying marketing stunt
both of those companies combined are smaller than the number of people microsoft is trying to buy out right now
[dead]
Anthropic 10x'd its employees over the past 3 years. Shouldn't it be cutting employees if that's the case? Especially because they probably have access to more powerful models which aren't released to the public yet, probably with fewer restrictions, and don't need to pay the markup for using their own service?
10x in three years for a "trillion dollar" startup is nothing
Not when the product you're selling is to make humans obsolete. Imagine if Apple had a budget to buy pixel phones for every employees.
I will keep assuming any news coming from AI companies are a type of corporate propaganda, and that AI companies frequently use their millions of dollars of marketing money to create a narrative to maximize hype and make money.
> I will keep assuming any news coming from AI companies are a type of corporate propaganda
That's just reverse naivity.
Reverse naivity? Never seen that term before, is that synonymous with healthy skepticism?
Recursive self-improvement is just around the corner. Any day now they'll be laying off all their employees as their god-like AI makes investors trillions of dollars. /s
You shouldn't be getting downvoted, people should be concerned that a company that's about to hit the valuation that it's at is not exploding in junior hiring and expanding their workforce like prior companies
Retail has shown no indication they won't happily foot the bill.