Opus 5.5 came out right at the end of the quarter, September 22. I'm confident that it caused more usage growth since then. I'm sure there are a lot of people rooting for their IPO to fail but they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
> they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
I think the first implying the second might be a bit of a stretch. At some point, companies are going to start looking at their AI spending and start negotiating deals. OpenAI is right behind them and Anthropic is around 8x more expensive (Sol 6.1 vs Opus 5.5 based on Artificial Intelligence's cost per task; cost per token is a somewhat useless metric when different models use vastly different numbers of tokens for the same thing). Chinese models like Deepseek 4.1 Flash are cheap, can be self-hosted, and while certainly behind Claude Opus 5.5, are still good.
Lots of companies buy "inferior" products like Microsoft Teams. Over the next 5 years, it seems likely that many models will get really good to the point that people often won't be able to tell any difference. Even if something isn't "definitively" better, companies might say "our engineers can't tell the difference and we're cutting our costs by 85%."
It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
It's not like it'd be crazy hard to analyze workers frustration as they did their jobs and figure out if one model was actually better. They could log all the conversations and it's not like AI can't easily determine things like "no, that's completely wrong" as a phrase showing that the agent was really off base.
I think a big part of the next 5-10 years is going to be about cost/efficiency.
Deepseek 4.1 flash is so fast and cheap. For coding tasks, it only seems to lack the robustness of Opus 5.5 rather than 'intelligence' - it fails to double check critical assumptions, stops before the task is over, and doesn't do enough speculation or exploration on its own initiative. Yet it can do all these things when prompted.
There seems to be another magic ingredient for improving agents beyond model size & efficiency, and so far only OpenAI, Anthropic, and maybe Google know what it is.
Ultimately the question is whether there's a pareto optimal point for intelligence models. There's 2 potential worlds:
1. Frontier models are so good that the token cost of an average task becomes miniscule, so users are incentivized to use frontier models because their sheer intelligence makes them cheaper to use.
2. There is a pareto optimal point where even if you spend more on tokens overall profits decrease because token costs rise faster than revenue does.
True, but I also believe that one can generally hire "more intelligence" if one is willing to pay more in salary and benefits, which most companies aren't. Basically, the original argument wr Microsoft Teams.
> It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
I’m not sure it’s a binary thing that contradicts.
I want AI and these companies to succeed. And I especially want new companies to succeed. But I want them to be valued far more realistically. Because if they have to be worth trillions, then there’s a real motivation to do what it takes to realize that valuation. And that’s how you birth a Zuckerberg.
I’m not rooting for them to fail. I’m rooting for the correction to happen sooner than later.
> But I want them to be valued far more realistically.
They are. That's the problem. The obligations that these companies are, on paper, supposed to meet require these valuations.
If you value Anthropic at $500b that's still a fuckload and vastly less than their obligations and a lot of people are going to say "well wait a minute how am I going to get paid".
The amount of spend going on, and planned, is insane and underpins a lot of this, and is why people are skeptical of the valuations. It has nothing to do with "Wanting" AI to fail and just a reasonable bet based on the math. For them to pay back what they have on their balance sheet entire continents will functionally be a spec in comparison, and that's really really fucking unlikely to happen.
Doubly so when you consider that there's a question of if they can be profitable to projected needed levels without fundamentally breaking some economics parts, and that this looks a lot like the overbuild of the internet during the dotcom bubble, which was also hideously overvalued at the time, for the time frame they were looking at, despite being THE literal world changing tech.
> If you value Anthropic at $500b that's still a fuckload and vastly less than their obligations and a lot of people are going to say "well wait a minute how am I going to get paid"
Not how valuations work!
Walmart made $187bn in sales last quarter of which $6.5bn became net income [1]. That supports $115bn in liabilities and a trillion-dollar market cap at 46x earnings [2].
If you want a more-extreme example, look at JPMorgan: Its $900bn valuation (14x earnings) [3] supports $4.6 trillion of liabilities and $5 trillion in assets [4].
Comparing valuation to liabilities to judge solvency is bunk on multiple levels.
> All established companies that have been net positive
On what metric? I just showed you JPMorgan has an equity value a quarter of its liabilities. JPMorgan is not insolvent.
Comparing liabilities to market cap doesn't tell you about solvency per se, it tells you about leverage. It tells you whether a solvency problem could be solved by selling equity. Not if there is a solvency problem.
> "it owes a bunch of people money, so it should be worth more"
"it owes a bunch of people money so the only way for it to not be insolvent is it must be worth X, and if we're talking lower than X this is already a problem"
> For them to pay back what they have on their balance sheet entire continents will functionally be a spec in comparison
Walk me through this math. I’m seeing $900bn Capex over a decade. That’s $90bn required free cash flow a year, on average. If they can do 25% margins, that requires $360bn annual steady-state revenues. That doesn’t require continent-scale anything.
This is the assumption we disagree on. They're net negative by an order of magnitude on most of their current customers. I don't see them flipping to break even + 25%.
Further you're assuming they can turn that on today, when they're still very clearly not ready. Every year the timeline gets worse.
Take them to 2% margin with 8 years to go and you're looking at needing 4.5t annual.
This is normally where I'd say "but 2% feels extreme..." and hedge in the middle, but like...the current pricing isn't even close to 25% margin, it's not even close to break even.
They're net negative by an order of magnitude on most of their current customers.
From the S-1, in the first 3 quarters of 2026 their revenue was $33.2B and compute costs were $18.8B. Inference is highly profitable, they "just" need more of it.
> Those numbers are stupid large for a company whose only moat is never ever being in second place
Not disputing that. Just saying you don't need continental-scale anything to hit these figures, you need annual revenues on the scale of what Walmart does in a quarter.
Anthropic has a near-clone in OpenAI (and vice versa). There are multiple smaller competitors, several much cheaper (especially the Chinese models).
Anthropic obviously needs continent-scale balance sheets, I would argue "multiple continent-scale balance sheets", since Google & co make their money by selling everywhere except for China.
Anthropic and OpenAI are never entering China on a massive scale. US relations with Europe are deteriorating and LLMs are <<massive> risks in an adversarial relation so it's not inconceivable Europe at some point starts marginalizing US tech companies, at least LLM labs. I imagine India will want to do something similar at some point.
Where exactly is Anthropic going to make all that steady revenue from, if it can't break out of coding and cybersecurity? Even worse, both those markets are super tech savvy so those cheap rivals are even more dangerous. One slip up the next 5+ years and Anthropic is history.
> There are 13 companies in <<total>>, on the <<entire planet>>, that make more than that
Would you describe UnitedHealth as a company that render "entire continents...a spec [sic] in comparison"? Sinopec or Petro China?
I'm not saying those growth requirements are obvious or even feasible. Just that they don't require some paradigm shift of global economics to achieve.
> Most of those needed decades to reach this point
How long did they take to reach Anthropic's current scale?
shareholder value of a company = equity = assets - liabilities.
Requiring equity > liabilities means assets > 2* liabilities.
That is nice, but not at all required for financial health. Especially for infrastructure heavy companies where the debt is used primarily to fund infrastructure buildout.
I can't disagree enough. I've gone from filling 4 Ultra accounts every week with Fable to barely managing to fill 50% of one with Opus. It's so damn efficient I'm spending 1/4 of what I was before with Anthropic.
For me, that's amazing, but maybe it explains these numbers and maybe it'll finally change when Fable 5.5 forces me back into high usage territory. If I even really need it's abilities as more than an advisor (words I didn't think I'd imagine myself saying back when Fable 5 was dropping).
As a counterpoint I haven't seen opus doing that much better than the latest gpt models? Astra is definitely comparable to fable and sol has been as good as opus for me. So this rings a bit anecdotal and that's probably what people will be seeing affect the growth trajectory too.
IMO Opus 5.5 has quietly been a revelation. It basically requires no hand holding from me given a sufficiently articulated prompt. At this point I'm limited by reviews and end-to-end validation of changes.
Are these two statements not at odds with each other? You can always say a prompt is not sufficiently articulated if you need to add additional handholding and vice versa, thereby invalidating any meaning of measurement.
What I mean is just a modest effort to guide the model up front is adequate for the full end to end session -- one-shotting, basically. I used to go through planning mode prior to 5.5, to make sure it understood my intentions but I no longer need to. I just ramble off the desired outcome and some broad architectural guidelines into wisprflow, hit enter, answer any questions it has and then a PR that tends to only get minor feedback is produced. It's basically fire and forget.
then a PR that tends to only get minor feedback is produced
How do you determine that this is because "Opus is so great" vs. "people are no longer reviewers but just approvers"?
As in, many if not most reviewers have always been nothing more than approvers, i.e. people that click the approve button with minimal to zero review, maybe a "pro forma" comment so nobody has them show up in some statistic as "only clicks the approve button".
With the amount of AI generated PRs this will have tended to become more of a problem, not less.
With earlier models simply describing your desired end result often went poorly, and instead you had to break your prompt down into incremental steps. A single “sufficiently articulated“ prompt would result in the model forgetting about some of the things you told it to do by the time it got there.
No, because "sufficiently articulated" is relative to the model. Opus 5.5, often is "please handle this ticket someone assigned me, check this link it might be useful..."
I’ve felt that way since the 4.1 releases. Opus 5.5 is a good model, especially after the disaster at the 4.7 mark.
But sol 6.1 is at least in the conversation as being as good and is much cheaper. I farm out bunches of my tasks to Luna and Haiku these days and have a working prototype of a glm 5.3 workflow at work that shows it being competitive.
The price pressure is very real and if you aren’t on a subscription then there is virtually no reason to pay the Anthropic premium
Every time I try and do something with codex/astra/sol 6.1 coming from Claude Code I want to scream. I try it regularly to make sure I'm not getting to tunnel visioned on Claude but at the moment it's night and day.
Not due to capability per se, but writing style and polish is much worse.
It will leave really bad comments in the code, commit messages etc without strong explicit steering. Though I've basically moved away from having the model write any documentation at all.
It also tends to talk very circuitously and referencing things implicitly instead of just being direct and clear.
Certainly it's more capable in debugging, but I much prefer Astra over Sol now.
Of course, Astra has its own problems of code golfing everything. But at least it's very concise and direct
if you go on short form social media, Opus's ability to do agentic work in things like blender, css animation, etc. have spawned hundreds if not thousands of channels of people pumping out content with it. That and all the video game shenanigans (and now Anthropic is even modifying its system prompt to limit some of these things!)
There's definitely unique things that so far only Opus is capable of doing.
Opus 5.5 is amazing and has also dropped our bill dramatically. Our engineers are spending 1/3 as much as they were before and seem actually happier with the results.
Anthropic can remove all access to their LLMs and start replicating entire software companies. Anyone relying on open weight models is at risk of a security breach from even slightly better models. It is ironic when people think Anthropic is the one without moat while they give all their data away.
There's a stock market truism that anyone that has real alpha doesn't sell it or give it away. I feel like there's an analogue here, that if it was just that easy for them to pack up their toys and go conquer the world, I can't imagine they would have been in the business of selling tokens at this point
IMO if llms keep improving, there's just not going to be a business in most of software anymore. It wouldn't matter if Anthropic tried to replicate all the saas companies because the category is soon to be dead. Why would you pay for someone else's software when you can prompt your own to your tastes? Every bit of software would just have one user
Yeah there was a vibe shift towards OpenAI being much more competitive for daily use and then Opus 5.5 hit. Will be curious to look back and see if that shows in data. Of course, could change tomorrow.
I expect growth to go down. Opus 5.5 is not only better, but a lot cheaper. My company was spending quite a bit on overages with Fable 5. Now we are no longer doing that.
Opus 5.5 may be occasionally better than other models on code (I find this debatable though, because Astra is very good and will often do better imo), but I would wager there's way more money to be made on things like image and multimodal tasks or at least nearly as much, and frankly even Gemini 3.8 flash can spank Anthropics best models on visual tasks and be absurdly cheaper; that also ignores the increasing aggressive classifiers Dario seems to like which block a lot of stuff and I've seen drive a number of zai subscriptions because blue teaming is increasingly very important.
This is likely true, but doesn’t really tell us much about future revenue growth.
Now they need to compete on price, I think their revenue growth may start to look much more like typical Jevons paradox examples.
I.e as their usage grows exponentially their revenue grows linearly rather than exponentially as it was before.
I’m wondering whether Opus 5.5 would have caused growth in an absolute sense or just growth relative to competitors.
Anthropic’s market share is pretty high and there is not enough total currently addressed market for growth at the expense of a competitor to make a big long-term difference.
Well what I do know as someone who has a bird’s eye view into many enterprises is that many of them are starting to realize they’re bleeding money to Anthropic and are telling their teams to be more efficient with tokens.
I've been throwing opus 5.5 at problems I got stuck on with sonnet in the past and it has breezed through them all. It really reinforced how far the models have come.
> I'm sure there are a lot of people rooting for their IPO to fail but they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
Then why all the rush to IPO then?
The strongest chinese models from DeepSeek, Kimi and z.ai are already good enough to run for hours on most coding tasks cheaply.
For capital to win they need to get their money out of Anthropic before the bubble deflates. This means dumping the risk on retail. There is only a finite amount of money available to buy AI stock, and the SpaceX IPO sucked a lot up, so it's a fight between Anthropic and OpenAI who is going to get what's left. The loser is left without a chair to stand on when the music stops.
I have zero sense of this myself, am just curious, and not confident anyone really knows, but: I get that its a big deal for software developers when each new model comes out, because its perceived (correctly or not) as the next big boost for their productivity, but what about everyone else who is not a coder? Is there big reactions there to newer models? I don't know a single person who is not a engineer that would be able to tell me if "Astra" or "Opus" are Anthropic things or an OpenAI things respectively, much less when and what the big new version is for any of them. Does this play out for others? Or am I just in my own bubble for some reason?
What does average, non-computer Joe or Jane think when the next big model comes out? "My emails are going to be so much better now, I better stay subscribed"? "I can't wait to see how well it answers my questions"?
One more anecdote: People at our company are very excited for integration of existing industry specific tools and processes.
Some of them do require frontier models but many can do with localish models. Some of these processes (in our industry) are also not really limited by software at all but by the speed heavy machinery can be operated and/or manufactured and/or transported.
I'm certain LLMs are going to continue rapid growth even there but it may be an ideal field for have lesser models work for longer if it saves money in the end.
Of course who knows in the end. I'm certain there is a high chance of this being entirely wrong
Anthropic S1 had numbers up till Q2. Many people speculated that Q3 numbers were lower because many companies had moved away from "token as measure of productivity" or tokenmaxxing. Anthropic Q3 numbers proved that theory.
As the article notes the growth is still huge. But it has slowed down relative to what they were growing.
We talk about revenue far later in a company’s lifecycle than we did before because high-tech, high-capex companies take a long time to become profitable.
Amazon is now the posterchild for this, but it happens in other industries like pharma a lot too. And in today’s stock market, where a fundamentally weak company like Tesla can trade at eye-watering multiples because of speculation, why worry about profitability at all? (Partly joking but that’s how people think.)
When and why we start expecting Anthropic to be profitable are great questions though.
Amazon has the world on lockdown though as the best place to buy stuff and get almost anything to my house in a day or so. I can use any AI model out there at ever decreasing razor thin margins and there is no moat or lock in at all. Nor do most people even need AI. I’ve certainly never and would never pay for AI anything. Anthropic is basically in the paper business, AI is a commodity at best, but maybe worse, because people actually need paper.
Amazon was unprofitable and not at all dominant for many years before reaching its current status. Books and groceries are also a race to the bottom of margins.
Up 47% quarter over quarter when the previous quarter was up 145% quarter over quarter and close to that the quarter before. "Growth flattened" is generous.
From what I could tell, it probably even distracted from the metrics that traditionally matter. I seem to remember tokenmaxxing dashboards being built, and other "shovels."
Completely anecdotal but I did find the better engineers tended to take to the AI tools quickest. At the beginning, tokens was probably a good way to pick them out. “Good” being entirely subjective around a product focus/mindset. If you ever find a good metric like that, it is critically important to never tell anyone, else it will quickly become not good.
> I don't think the "tokens consumed" metric was a very good one.
When mid or high level tech leadership pushes from top-down new things, that look dumb to the engineers, the leadership is either stupid or there are hidden incentives that regular people don't see. In my experience mid level usually matches with the former, and top level with the latter.
Interesting angle to blame the slopmaxers when the AI companies have been pushing maximum usage for a while now.
I assume they’d hoped that the value of the spend would manifest itself and they could raise prices. The opposite happened and companies are instead cutting back.
It sounds to me like they chose a poor marketing and pricing strategy for a product that is very difficult to make cost effective.
Everyone trying to vibecode their next million dollar idea only discover the web is dead. Nobody is looking for anything because that would require leaving the safety of the same handful of sites everyone else is on.
Because not only is there no guarantee that the peddled slop will be online or maintained or supported for any amount of time after its debut, but it's safer and more effective to just reap the slop as context into something slightly less sloppy.
It may have something to do with the summer intern season ending. I know one of the interns hired by my employer used up the allotted AI spend by doing “fun” things unrelated to the internship project, like making Claude write an entire SQL-to-Pandas transpiler in 10,000 lines of Python. I can’t blame them for using it extensively when given access to a generous limit on token usage.
When was the last time anyone used the term "tokenmaxxing" right now?
Exactly, No-one and as predicted it was a fad. [0]
Vibe-coders don't want to risk getting their accounts suspended if you use their models in a certain way.
We are seeing the result of those subsidizations and discounts that Anthropic made this year and how much of this "tokenmaxxing" hype was used to make their revenue numbers look "good".
With the arrival of more cheaper models which are good enough, it doesn't make much sense to use Anthropic's models.
So don't be easily fooled by their revenue projections. Rather question if it is sustainable for them. (Spoiler, it isn't)
Lmao a colleague who works elsewhere said some goon at work created an AI leaderboard. He spun up research agents. Top of board achieved. Granted the guys a very solid developer, but he did it to prove how dumb this is.
At the beginning of coding agents really picking up adoption (I guess 12-18 months ago now or so) I remember this sort of thing being so common that some startups were trying to sell a turnkey solution for "leaderboards".
I'm not aware of any of those products really sticking around.
We can finally dispel the myth that management is competent when they, almost as one across industries, tied continued employment to setting fire to company money
AI put a massive spotlight on all the required but unspoken decision making and coordination that must occur with stakeholders outside their own teams during the development process.
Dysfunctional middle management has always traditionally neglected it or actively undermined it, blaming others when things go wrong and taking credit when things go right.
Naive developers are always more than happy to do what they think is "picking up the slack" not realizing it's a trap.
Perhaps most organizations are just now more wise to the fact that using AI to "accelerate the process" is like arming the monkeys with machine guns?
It's a combination of natural language being an awful programming language and frontier ai companies optimizing for longest time horizon at 50% success. I mean what did they expect? Length of time to not check in and still bumble around to turn a single prompt into a literal half-assed solution.
Your machinery was not written elegantly. It was sloppily assembled via evolutionary pressures.
A gene gets duplicated, mutates, and suddenly you have new functionality and new fitness criteria.
Try something bad and you die without reproduction.
Our code isn't clean. There's redundancy and extra packing and slop all over the human body.
Embrace it. Stochastically hill climb.
Code is DNA. Evolve it, don't intelligently design it. Fit the curves of your environment quickly and use tests to cull bad behavior and the slopes of your domain.
I know this is tech sector funny money and we do just disregard profit and judge companies on revenue alone, but the 5.5 models that just released are quite a bit cheaper than the 5.0 tier.
I wouldn't really be surprised if better model efficiency was part of the reason for a revenue stall, but I would also expect their costs to have dropped a fair bit too (although I have no idea whether that corresponds to an actual cost reduction or if it's all tied up in hardware and land that they already bought and will need to hit more utilisation of now.)
I think most people who care or are tracking it would say we've hit it, and those who disagree just have their goalposts a lot more clearly defined and are quietly waiting.
The very first goal post for AGI must be that it can act, learn, and grow independently. They must be self healing, self correcting, and anti fragile. That's definitely what the G in AGI is about. But all models fall into decoherence if you let them work on their own too long.
If you left an AI alone for a long time and when you came back they had built up a working society and learned new useful things, then that's the proof it's AGI. From this definition you only know it's AGI after you let it loose.
I don't understand how anyone can call these AGI. I've always had the view that AGI == Independent human-level intelligence.
At some point someone moved the goalposts for this to include LLMs that emphatically are not that. I think it was Altman when they came out with Astra?
they aren't as intelligent as a human, but they can throw shit at the wall faster, and get something that sticks faster. they are superhuman in areas that require trial and error, like hacking.
everyone stopped talking about the Turing test once ChatGPT came out and broke right through it while still being very very dumb. AGI is just too blurry to define.
If the hypothesis is that "everyone is tightening spend on usage", then anthropic is boned, if that holds true. Look at the price to performance for any benchmark and compare OAI and Anthropic, and this would be a pretty clearly terrible reality for Anthropic. OpenAI absolutely blows Anthropic out of the water when it comes to cost, if you pay per token.
> The real question is how many people haven't figured out how to use AI to make their job easier/more productive.
I think the answer, to a reasonable approximation, is “everyone”. Outside software engineering most people aren’t using it for anything useful, and I keep encountering technical people including programmers who don’t use it either, which never fails to surprise me.
One of my projects is a transportation management system (TMS) and of course I have added a lot of features to it including a very capable chat-based UI that can automate a lot of tedious data entry and other tasks for dispatchers and accounting and so on. It gets used but really not very much. I’m starting to realize that I’ve probably taken updates as far as they can go in terms of the ability of people to absorb change. Most people just carry on doing things the way they always have.
We had one person on the admin team who started using Cowork to automate all kinds of stuff. She loved it. Then she left. It hasn’t caught on with anyone else.
This is not to say that things won’t change but I think it will take a while.
my brother uses it to do ocr on his books but, lately, claude has been refusing because of copyright or whatever so i've had to do it for him (using ai, yes, but not anthropic's)
I know firefighters and machinists heavily using it, the only places I see slower adoption is purely because of bad leadership- exclusivity contracts with certain providers or just plain fear.
I think this dynamic is underappreciated. Another thing I see in working with different companies (consulting) and discussing with colleagues at different jobs is that we as humans tend to do a pretty bad job of modeling business outcomes and the dynamics that govern them. There are lots of outcomes where you can identify one contributing role, increase the productivity for that role by a significant factor, and still ultimately not move the needle much, because the constraints span multiple functions.
As an example outside of software engineering, a lot of people in marketing roles have found agents to be incredibly useful for particular tasks. I've talked to multiple people who wouldn't standardly be classified as "technical hires" who are using Claude to manage things like digital advertising. It's valuable here because managing campaigns on a platform like Google or LinkedIn involves a ton of granular tweaking and analysis. Every performance marketer makes tradeoffs around what they're realistically able to spend time doing, and what makes the biggest impact. So if you have 250 campaigns that you manage solo, you likely don't have the time to really manually optimize the bidding and budget for each campaign every couple of days based on new data. You can only try so many new angles, because creating new media is time consuming. You don't have time to construct custom experiences on-page for every particular campaign. There's a whole industry of tools that turn a tidy profit automating parts of this for this reason. And agents are really good at it, because a lot of it is just applying the same analysis over and over again.
But for a lot of startups, increasing traffic from ads 20% while decreasing spend is a good outcome, but it doesn't actually turn into revenue. The product itself might just be limited, and already being discovered by the narrow segment of potential users who find it valuable enough to pay. Or the sales process might be kind of busted, meaning the team will really only close a certain volume of new deals in a given month, no matter how much inbound they have. Or some combination of all of the above.
If it's just one marketer using an agent to optimize their advertising, there's not a lot they can do about the other constraints. And so what you end up with is one employee who feels really productive and sings the praises of Claude, but not a meaningful lift in the higher level outcomes you care about, and the analysis gets messy. In an ideal world, everyone realizes where the other bottlenecks are, but that's a genuinely hard thing to do and people have complicating incentives.
I don't think this is a particularly new problem. It feels identical to any other situation where a department asks for a tool, loves it, and it doesn't really make a difference to the bottom line or make anyone else's life easier in noticeable ways.
> very capable chat-based UI that can automate a lot of tedious data entry
Because rewieving chat bots data entry is even more tedious then doing it yourself. And as long as they have to review it, it does not even matter that it makes only few mistakes.
I see it with JIRA and Rovo. It just sux using it, despite it being reasonably competent at creating tables and what not. If I could do markdown as used to be normal, it would be appreciated a lot more.
We are capped by testing and PR review. The junior devs can spam out a million PRs a day but we keep having production incidents from not carefully scrutinising the changes. So the speed of code generation is limited by how fast we can actually accept those diffs.
Production incidents from not carefully scrutinizing changes is a sign that someone more senior should be putting infrastructure in place to minimize the impact of those incidents.
Feature flags, staggered rollouts, quick rollbacks, smoke testing, etc.
Sure, if you were hand managing every rollout up until now, and relied primarily on unit tests, you’re in for a world of hurt when you give a junior engineer Claude code and let em rip.
But if you have more advanced systems to measure and maintain overall service health, and codify your business logic invariants, you can really start flying these days as long as you’re all aligned on what to build and the potential impact.
I agree, with a big tech deployment process this would be less impactful. But getting from a company used to manual testing and weekly deployments to a fully automated CD pipeline is more work than a single Claude prompt.
Unironically you should be using claude for monitoring and within CI. "PRs causing bottleneck" is an anti pattern during AI. It is as easy to verify as it is to generate.
That was always the issue... Rewrites were done before LLMs and the results were the same... There was never a point in time where the bottleneck was production of code, it was always review, testing, QA, fixing bugs and edge cases, and that is when only focusing on the technical part of delivery, and leaving out communication and synchronisation within the company and between company and clients...
That strongly depends on context. In larger organisations, this is definitely the case. For smaller startups and teams then the overhead is much smaller and the actual core programming work is much more of a bottleneck.
Sometime in the last 15 years we got these huge backlog wishlists of one line thoughts. People were convinced because software devs couldn't keep up translating one line thoughts into working software that the code must be the bottleneck
They never realized how much effort went into turning those one like tickets into code was, and the code was the least part of the effort
How so? Or is this just vibes based analysis? I dont really know what you mean by classical engineering, but sure if something can be done on a computer it can infinite monkey theorem its way to a solution. This isnt exactly useful its just an enablement of laziness.
Idk how many people need to hear this, but not everything 'is computa',no matter what Trump says.
Much of the worlds problems are not verifiable with a compiler that can be trained or brute forced against.
Llms are not as useful as ya'll think they are, not everything exists in a web browser, which is why Anthropic et al have saturated their TAM and are now frantically hiring sales people to convince managers in other verticles.
Right, this is what astonishes me about the constant rhetoric/hyperbole about how this is going to impact "all the jobs" and make everyone wildly more productive than ever before. It's a fabulous tool in a limited number of areas, and it's otherwise only handy here and there. But just because it's a fabulous tool for the coding that is so prevalent among HN users does not mean that it's a fabulous tool for everything.
A hammer is fantastic for driving nails, but it's not really all that good at washing a dog.
This is conflating the benefits found within those areas with the ability of AI to actually deliver those things effectively across every single profession. Improving something by 2% vs. revolutionizing an entire industry are wholly different propositions. Maslow's hammer would like a word.
Mechanic, welder, builder, farmer, logger, millwright, trucker, heavy equipment operator... I could go on. Literally the entire economy in the area I live is completely impenetrable to tech, let alone AI. None of it is worth a shit to any business here.
These are all jobs the silly con valley idiots claim continuously to be automating away, but it's all just fuckin lies.
The extent of tech's penetration into this economy is internet ads on facebook and streaming services. Nothing to do with actual work.
I use classic engineering to refer to the (engineering) disciplines where mother nature messes your day up rather than obscene levels of man made (or machine made) pure complexity ruining your day.
It's easy to understand how LLMs got so good at coding, they can rapidly iterate, self test, and it all happens in their digital universe.
These models now starting to spit grey beard knowledge of precision analog electronics, lightweight fastener design, and chemical surfactant systems is something very very recent. A few years ago I think we were pretty confident that the tech would hit a wall at the digital divide. Well, I can tell you now that it did not. It seems it kinda blasted right past it.
AI growth is unlikely to be linear. We have seen a lot of hype in the last 12-18 month that give us ideas like "tokenmaxxing". But this is the same cycle that we have been before: hype -> disillusionment -> someone figures out a better way to use it -> next hype phase starts.
As it stands, even the people closest to this technology have barely figured out how to use it effectively.
AI has historically had a habit of disappearing for ten years and coming back once memories fade and the new story sounds enough different from the old one.
I don’t think this era is going to end with us all unemployed, but I do plan to be retired before the next cycle.
For big ones it’s often network effects and/or friction. Those can be pretty good moats even with LLMs. These models are also nowhere near the point where they can actually reverse engineer complex products.
Network effects have subsequently been; shown to be logarithmic. You’re not getting 100x the benefit from 10x the users. You’re lucky to be getting 12x the benefit and you also have logarithmic increases in costs because there’s no such thing as economies of scale in software. You pay for every order of magnitude in hardware or software development costs.
It was six months last year, it’s on the order of weeks now. Presumably they have unreleased models that are a step ahead still. The moat is capacity and how much of an ecosystem the can build around the llms; eg anthropic figma tool.
How much easier my life would be if Chinese open models were even just 1 year behind outside of X posts and one-shots.
My easy litmus test is would the year old model (Opus 4.5) being open sourced today would eat the traffic of every single Chinese model... and the answer is pretty much yes.
Every Chinese model would throw away their work just to post-train or distill Opus 4.5 and get a more reliable model.
I think the next cycle of growth will be from increased subscription pricing? Thats what the streaming companies did - start low, keep increasing forever. OpenAI already merging codex / chat and I think most people will have to buy that min subscription or even the higher tier. I am already at the $100/month
I am not saying its a must, but those who are reaping the benefits in someway will end up spending more I guess?
This is naive. AI is one of those few things that does have moat. Its shockingly hard to hire people and create a good company culture that does frontier training. There's only like 3-4 labs in the world and the rest are distilling.
Even in next 3-4 years, the company that has the best ecosystem of researchers, culture, shared knowledge and processes will be able to churn out models in a way that can't be replicated.
i think its a lot deeper than that, compared to deepseek-4.1-flash, glm-5.3-flash and the recent minimax-m3.1-flash [0], the comparison in price (15x cheaper inference cost) and the productivity gains are very similar if not more profound depending your prompt topic (cyber, biochem, etc)
I don't see the issue. If the market is saturated by them, the slowdown seems to indicate to me that it has reached (at least transiently) its intended audience. If the end goal is that every person on earth has a claude subscription (or whatever your choice is) then yeah that is worrying, but I don't deem that realistic.
Why do you say this? If they get to even 100B RRR they will pay off all of their commitments in <10 years with money to spare. Does that seem so impossible to you?
Flattened might be an exaggeration but they lost many a customer with boneheaded decisions like blocking subscription access to OpenCode. I don’t think they’re past the point in their growth curve where tinkerers don’t matter anymore. Again, definitely not flattened but not as parabolic as they could be.
He says a lot of things, but one specific thing I remember from weeks ago is that he said Anthropic growth slowed on Q3, and that's why they decided to push for the IPO (and use only Q2 numbers).
It reminds me of the Apple TV rumors. For ~4 years the new AppleTV was just around about to be announced for <reasons> and you should wait to buy one. Then it's finally announced and people are all "yeah, but a few weeks before the AppleTV was launched they said it was going to have more RAM!".
Sure, and they said similar kinds of things for 4 years. Why act like the broken clock is suddenly worth noting just because the time was finally right this moment?
my simplified take is the headwind is on two ends - smaller models being sufficient enough & open weight models then second utility of 'a.i' is maximal when deployed with humans i.e engineers applying a.i on top of current applications putting an agent layer on top.
the second part is not sexy - but we see all a.i labs having fde arms e.g ode, microsoft frontier company etc but then again that's the part you can't hype.
The conclusion that this is a bubble seems arbitrary and disconnected from the several true facts outlined in the article. Anthropic's growth _is_ slowing, albeit still strong at ~47%. They do have enormous compute obligations, and it will require a lot of growth to meet them. However, they also increased revenue per dollar spent on compute from $0.40 to $2.08 in five quarters and hold $131.7bn in cash.
Keep in mind this growth has mostly been sustained by low-level business automations, coding, and chatbots. With the adoption of products like personal agents, e.g., Meta's Muse, I am pretty optimistic there will be demand and growth around this level will be sustained for a few quarters. Doubtful we will ever see another ~150% growth quarter.
> revenue per dollar spent on compute from $0.40 to $2.08
So that’s ignoring labor and perhaps facilities? That’s like saying we’ve hit net positive fusion by only considering the power produced by the laser and ignoring the 80% power losses in the feed-in.
I am actively pushing my workplace to scale back our AI usage (claude cowork)... the amount of AI slop being pushed between teams is off the charts and ends up creating a mountain of work for reviewers. Hurts productivity rather than improves it.
- Would you like to get paid a million dollars right now, or get paid a penny today that doubles everyday for 30 days.
- hmmm, why 30 days? Make it 60 days and I'll take the penny.
The effort needed to double everyday, or month is realistic at first. But then you can't just keep doubling without swallowing the universe along the way.
You could in theory be responsible for the destruction of the entire solar system in order to pay you. Sit on your throne atop a mountain of skulls and weep.
Right, but a comparison is always one more implicit derivative. If you look at revenue itself, it's growing or shrinking, that's first derivative. So there's one more derivative here than usual?
I guess the interesting thing would be to look at actual company assets/value rather than revenue; revenue already in itself is a derivative. I.e., contrast their revenue against expenses. Last I heard those numbers are deep red, aren't they?
Growth is first derivative. Deceleration is second derivative. Focusing on second not first I think was the point, not implying there is some nth derivative under discussion.
It's because everyone is talking about how the bubble will burst! When people talk about it a lot, they tend to believe it.
Everyone just needed to give their computers to Sam so that no one could talk about the bubble, and then everything would be fine.
If you ever wonder why Anthropic is going all out to try and fearmonger their way to an open weights ban, this graph will provide all the answer you need.
Their financial statements suggest the answer, even when you take into account the cost of training the model, is now 'not at all'. They are profitable, the main questions for their valuation are 1) can they keep their margins, 2) how big can they grow?, and 3) can this growth match what they have committed to?
Question one is how do their investors get paid back. I found Series E through Series H totaling $111.5 billion in investments. Since series E was $3.5 I’m assuming all the previous rounds added up to less than that, so let’s say 115 billion at the outside.
That debt will transfer to public shareholders at the IPO but also will double or triple in the process, as the investors expect to be paid very, very well. So the current shareholders are looking to extract something over $400 billion from the IPO.
With the current revenue they might pull off that P/E. But they’ll still be extracting some of that value from the customer base, because shareholders will demand it.
Not at all, the subscriptions have an adjusted gross margin of approximately 85%. Very profitable! The training is a different story though. How long will the subs last if they stop training? Also, why on earth would they put on the breaks and try to be profitable when they have plenty of funds and are in the growth stage?
The subscriptions are a key part of the engine to ensure The Machine can keep improving itself by harvesting all that data and feeding it back in. Many of the enterprise contracts forbid the lab from training on their data, so the subscriptions are a form of gold.
Ah the beauty of circular financing. By owning the company who pays me to provide a service, I stand to gain massively by giving them a discount, making them look profitable, and fooling investors in to providing my exit liquidity.
The criticism has been about folks downstream of Anthropic being circularly financed. I haven’t seen credibly claims of Anthropic’s revenue being circular, but if you have a source, by all means.
I've been saying it for a while now but if someone sold me a box with model perf from ... I guess November 2025 then that's decent enough for my usage. Tool interaction is the magic, I don't need much if any more innate intelligence. Things are fast enough.
I wonder if Anthropic will "see the light" and stop trying to burn a bunch of money and sell a thing they already have (which is decently useful). Very ready to be proven wrong in any direction
wow really? Nov 2025 I was still fighting blatant hallucinations in helm charts. Now I'm two prompts away from a new build published on the app store and 5 video reels in the queue complete with captions and everything
The thing is I'm not spinning up a new app every day, I'm just working on a large existing system with enough constraints. From that perspective tools + the general rigor I already try to apply means that I can get some value out of LLMs without waiting for it to do something "good".
"I can one-shot a new thing" is interesting but definitionally a thing is new only once. Maybe if I worked in a space where we had way more throwaway things to work on.
At the end of the day I want to understand the output anyways, so it one shotting a thing vs me going back and forth with it in smaller increments is a mixed bag anyways.
I think one thing that could explain it is that some of the big tech companies might have been big customers of theirs, and now they have alternatives that, according to recent news, their employees are being pushed to use instead. Some of those reports mentioned internal usage getting cut in half. If this is true, it doesn't necessarily mean their growth has stalled.
Employees were and are pushing for more model and harness diversity too I promise.
But perhaps for different reasons than Microsoft, Meta, and maybe Adobe leadership are now. See, if they are all in on Claude how do they sue Anthropic for assisting in copyright violations and IP theft? The services would be cut immediately.
Well when they don't need them anymore..
This is why Anthropic is heaping on more and more model restrictions running up to IPO. I believe it's a smokescreen for the IP risk they outlined in their IPO filing..
1) customer concentration risk
2) when those customers use a diverse set of models - so too will the rest - they’re not idiots who want to pay more than they should
Aka % growth rate decreasing sharply and eventually flattening out. Within 2-3 quarters.
This is why Musk rushed consolidating xAI with Spacex and rushed to an IPO. And, why it was so important to him that Nasdaq and Russell rewrite their rules so he could dip into everyone's retirement accounts as quickly as possible. I don't think he's particularly bright, but he has an incredible cunning about how markets behave and how to manipulate those markets to his ends.
The writing was already on the wall about a lot of things WRT the AI market: 1. there's a limit to how much money will be thrown at it before investors expect a return, and that limit is fast approaching, 2. it's turning out "intelligence" is going to be a commodity and it's not much of a moat, 3. cornering the market on memory and GPUs (even if they sit in a warehouse for months or years before going online, as long as it drives up the price for competitors) only works if your competitors don't figure out how to do it more efficiently and on alternative platforms.
Spacex also soaked up a lot of the retail appetite for AI, even though it's an also-ran that barely qualifies, in terms of enterprise and government usage (and for good reasons, both technical and otherwise). Every other AI IPO is going to be facing a more pessimistic market, especially if we continue to be in a tough economic environment.
In short: They waited too long, now they're in a world where they've burned so much money people expect them to deliver something that no one in the AI industry probably can deliver: a monopoly-like hold on a huge percentage of the market, and to invent a bunch of new markets that aren't currently buying AI. They have to have a dominant position for this to work because they have to be able to set their price higher than is standard now, without everyone fleeing to other cheaper providers. And, when the market for memory and GPUs finally gets back to normal and supply catches up with demand, I'd wager a lot of companies, and even individuals, are going to get real serious about open models and self-hosting (unless prices remain competitive for the hosted models).
> 3. cornering the market on memory and GPUs (even if they sit in a warehouse for months or years before going online, as long as it drives up the price for competitors) only works if your competitors don't figure out how to do it more efficiently and on alternative platforms.
My understanding is that the GPU market is largely sold out for the next few years?
What exactly is the threat here? Like, which specific competitors, what efficiency improvements, what alternate platforms?
I didn't think I needed to spell it out for the HN crowd, since folks talk about the Chinese models every day. But, OK...
At least three Chinese companies have bragged about their latest models being trained and/or served on Chinese-made GPUs. Literally within the past 24 hours, LongCat spent a large portion of their release announcement for 2.5 bragging about how efficient their training process is and that it runs on domestic (Chinese) GPUs: "LongCat-2.5 uses long-sequence training during both mid-training and SFT. With less than 80 GB of memory per domestic AI accelerator, the training system prioritizes memory efficiency and aligns its parallelism with the supernode architecture."
DeepSeek has repeatedly blown the curve in terms of how efficiently a model can be trained and served, Xiaomi and Z.ai aren't far behind, and have also discussed training and serving on home-grown hardware. I'm sure you can search Arxiv for all of the papers from these providers (DeepSeek, in particular, who document seemingly everything, even when they're doing something ahead of the US providers) as well as I can.
Opus 5.5 came out right at the end of the quarter, September 22. I'm confident that it caused more usage growth since then. I'm sure there are a lot of people rooting for their IPO to fail but they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
> they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
I think the first implying the second might be a bit of a stretch. At some point, companies are going to start looking at their AI spending and start negotiating deals. OpenAI is right behind them and Anthropic is around 8x more expensive (Sol 6.1 vs Opus 5.5 based on Artificial Intelligence's cost per task; cost per token is a somewhat useless metric when different models use vastly different numbers of tokens for the same thing). Chinese models like Deepseek 4.1 Flash are cheap, can be self-hosted, and while certainly behind Claude Opus 5.5, are still good.
Lots of companies buy "inferior" products like Microsoft Teams. Over the next 5 years, it seems likely that many models will get really good to the point that people often won't be able to tell any difference. Even if something isn't "definitively" better, companies might say "our engineers can't tell the difference and we're cutting our costs by 85%."
It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
It's not like it'd be crazy hard to analyze workers frustration as they did their jobs and figure out if one model was actually better. They could log all the conversations and it's not like AI can't easily determine things like "no, that's completely wrong" as a phrase showing that the agent was really off base.
I think a big part of the next 5-10 years is going to be about cost/efficiency.
Deepseek 4.1 flash is so fast and cheap. For coding tasks, it only seems to lack the robustness of Opus 5.5 rather than 'intelligence' - it fails to double check critical assumptions, stops before the task is over, and doesn't do enough speculation or exploration on its own initiative. Yet it can do all these things when prompted.
There seems to be another magic ingredient for improving agents beyond model size & efficiency, and so far only OpenAI, Anthropic, and maybe Google know what it is.
Its not very complex to create an Agentic loop with deepseek and the performance increases a lot (and is still cheap).
What's your setup? Flash in opencode requires far more oversight for me than Opus 5.5 in claude code.
Seconded. Also curious to know.
Ultimately the question is whether there's a pareto optimal point for intelligence models. There's 2 potential worlds:
1. Frontier models are so good that the token cost of an average task becomes miniscule, so users are incentivized to use frontier models because their sheer intelligence makes them cheaper to use.
2. There is a pareto optimal point where even if you spend more on tokens overall profits decrease because token costs rise faster than revenue does.
The "no lab has a clear lead" world leads to 2.
You bring up a very good point. Not only do lots of people buy inferior products, lots of people employ inferior intelligences
To be fair, hiring intelligence is not easy.
True, but I also believe that one can generally hire "more intelligence" if one is willing to pay more in salary and benefits, which most companies aren't. Basically, the original argument wr Microsoft Teams.
> It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
That's a great product idea!
I’m not sure it’s a binary thing that contradicts.
I want AI and these companies to succeed. And I especially want new companies to succeed. But I want them to be valued far more realistically. Because if they have to be worth trillions, then there’s a real motivation to do what it takes to realize that valuation. And that’s how you birth a Zuckerberg.
I’m not rooting for them to fail. I’m rooting for the correction to happen sooner than later.
> But I want them to be valued far more realistically.
They are. That's the problem. The obligations that these companies are, on paper, supposed to meet require these valuations.
If you value Anthropic at $500b that's still a fuckload and vastly less than their obligations and a lot of people are going to say "well wait a minute how am I going to get paid".
The amount of spend going on, and planned, is insane and underpins a lot of this, and is why people are skeptical of the valuations. It has nothing to do with "Wanting" AI to fail and just a reasonable bet based on the math. For them to pay back what they have on their balance sheet entire continents will functionally be a spec in comparison, and that's really really fucking unlikely to happen.
Doubly so when you consider that there's a question of if they can be profitable to projected needed levels without fundamentally breaking some economics parts, and that this looks a lot like the overbuild of the internet during the dotcom bubble, which was also hideously overvalued at the time, for the time frame they were looking at, despite being THE literal world changing tech.
> If you value Anthropic at $500b that's still a fuckload and vastly less than their obligations and a lot of people are going to say "well wait a minute how am I going to get paid"
Not how valuations work!
Walmart made $187bn in sales last quarter of which $6.5bn became net income [1]. That supports $115bn in liabilities and a trillion-dollar market cap at 46x earnings [2].
If you want a more-extreme example, look at JPMorgan: Its $900bn valuation (14x earnings) [3] supports $4.6 trillion of liabilities and $5 trillion in assets [4].
Comparing valuation to liabilities to judge solvency is bunk on multiple levels.
[1] https://www.sec.gov/ix?doc=/Archives/edgar/data/0000104169/0...
[2] https://finance.yahoo.com/quote/WMT/
[3] https://finance.yahoo.com/quote/JPM/
[4] https://www.sec.gov/ix?doc=/Archives/edgar/data/0000019617/0...
All established companies that have been net positive. Not up and coming tech that's yet to come close.
> All established companies that have been net positive
On what metric? I just showed you JPMorgan has an equity value a quarter of its liabilities. JPMorgan is not insolvent.
Comparing liabilities to market cap doesn't tell you about solvency per se, it tells you about leverage. It tells you whether a solvency problem could be solved by selling equity. Not if there is a solvency problem.
> They are. That's the problem. The obligations that these companies are, on paper, supposed to meet require these valuations.
How does that make sense lol
"it owes a bunch of people money, so it should be worth more"
i don't...
> "it owes a bunch of people money, so it should be worth more"
"it owes a bunch of people money so the only way for it to not be insolvent is it must be worth X, and if we're talking lower than X this is already a problem"
The point of valuation is NOT to make sure investors waste as much money as necessary to turn insolvent company to solvent.
It is to figure out what ia its correct price. Otherwise the financial system turns into a scam.
> For them to pay back what they have on their balance sheet entire continents will functionally be a spec in comparison
Walk me through this math. I’m seeing $900bn Capex over a decade. That’s $90bn required free cash flow a year, on average. If they can do 25% margins, that requires $360bn annual steady-state revenues. That doesn’t require continent-scale anything.
> If they can do 25% margins,
This is the assumption we disagree on. They're net negative by an order of magnitude on most of their current customers. I don't see them flipping to break even + 25%.
Further you're assuming they can turn that on today, when they're still very clearly not ready. Every year the timeline gets worse.
Take them to 2% margin with 8 years to go and you're looking at needing 4.5t annual.
This is normally where I'd say "but 2% feels extreme..." and hedge in the middle, but like...the current pricing isn't even close to 25% margin, it's not even close to break even.
They're net negative by an order of magnitude on most of their current customers.
From the S-1, in the first 3 quarters of 2026 their revenue was $33.2B and compute costs were $18.8B. Inference is highly profitable, they "just" need more of it.
Those numbers are stupid large for a company whose only moat is never ever being in second place.
> Those numbers are stupid large for a company whose only moat is never ever being in second place
Not disputing that. Just saying you don't need continental-scale anything to hit these figures, you need annual revenues on the scale of what Walmart does in a quarter.
> That doesn’t require continent-scale anything.
Wha???
There are 13 companies in <<total>>, on the <<entire planet>>, that make more than that: https://companiesmarketcap.com/largest-companies-by-revenue/
Most of those needed decades to reach this point.
Anthropic has a near-clone in OpenAI (and vice versa). There are multiple smaller competitors, several much cheaper (especially the Chinese models).
Anthropic obviously needs continent-scale balance sheets, I would argue "multiple continent-scale balance sheets", since Google & co make their money by selling everywhere except for China.
Anthropic and OpenAI are never entering China on a massive scale. US relations with Europe are deteriorating and LLMs are <<massive> risks in an adversarial relation so it's not inconceivable Europe at some point starts marginalizing US tech companies, at least LLM labs. I imagine India will want to do something similar at some point.
Where exactly is Anthropic going to make all that steady revenue from, if it can't break out of coding and cybersecurity? Even worse, both those markets are super tech savvy so those cheap rivals are even more dangerous. One slip up the next 5+ years and Anthropic is history.
> There are 13 companies in <<total>>, on the <<entire planet>>, that make more than that
Would you describe UnitedHealth as a company that render "entire continents...a spec [sic] in comparison"? Sinopec or Petro China?
I'm not saying those growth requirements are obvious or even feasible. Just that they don't require some paradigm shift of global economics to achieve.
> Most of those needed decades to reach this point
How long did they take to reach Anthropic's current scale?
shareholder value of a company = equity = assets - liabilities.
Requiring equity > liabilities means assets > 2* liabilities.
That is nice, but not at all required for financial health. Especially for infrastructure heavy companies where the debt is used primarily to fund infrastructure buildout.
I can't disagree enough. I've gone from filling 4 Ultra accounts every week with Fable to barely managing to fill 50% of one with Opus. It's so damn efficient I'm spending 1/4 of what I was before with Anthropic.
For me, that's amazing, but maybe it explains these numbers and maybe it'll finally change when Fable 5.5 forces me back into high usage territory. If I even really need it's abilities as more than an advisor (words I didn't think I'd imagine myself saying back when Fable 5 was dropping).
As a counterpoint I haven't seen opus doing that much better than the latest gpt models? Astra is definitely comparable to fable and sol has been as good as opus for me. So this rings a bit anecdotal and that's probably what people will be seeing affect the growth trajectory too.
IMO Opus 5.5 has quietly been a revelation. It basically requires no hand holding from me given a sufficiently articulated prompt. At this point I'm limited by reviews and end-to-end validation of changes.
> It basically requires no hand holding from me
> given a sufficiently articulated prompt
Are these two statements not at odds with each other? You can always say a prompt is not sufficiently articulated if you need to add additional handholding and vice versa, thereby invalidating any meaning of measurement.
What I mean is just a modest effort to guide the model up front is adequate for the full end to end session -- one-shotting, basically. I used to go through planning mode prior to 5.5, to make sure it understood my intentions but I no longer need to. I just ramble off the desired outcome and some broad architectural guidelines into wisprflow, hit enter, answer any questions it has and then a PR that tends to only get minor feedback is produced. It's basically fire and forget.
As in, many if not most reviewers have always been nothing more than approvers, i.e. people that click the approve button with minimal to zero review, maybe a "pro forma" comment so nobody has them show up in some statistic as "only clicks the approve button".
With the amount of AI generated PRs this will have tended to become more of a problem, not less.
With earlier models simply describing your desired end result often went poorly, and instead you had to break your prompt down into incremental steps. A single “sufficiently articulated“ prompt would result in the model forgetting about some of the things you told it to do by the time it got there.
No, because "sufficiently articulated" is relative to the model. Opus 5.5, often is "please handle this ticket someone assigned me, check this link it might be useful..."
I’ve felt that way since the 4.1 releases. Opus 5.5 is a good model, especially after the disaster at the 4.7 mark.
But sol 6.1 is at least in the conversation as being as good and is much cheaper. I farm out bunches of my tasks to Luna and Haiku these days and have a working prototype of a glm 5.3 workflow at work that shows it being competitive.
The price pressure is very real and if you aren’t on a subscription then there is virtually no reason to pay the Anthropic premium
Every time I try and do something with codex/astra/sol 6.1 coming from Claude Code I want to scream. I try it regularly to make sure I'm not getting to tunnel visioned on Claude but at the moment it's night and day.
I've found Sol 6.1 a regression over 5.6.
Not due to capability per se, but writing style and polish is much worse.
It will leave really bad comments in the code, commit messages etc without strong explicit steering. Though I've basically moved away from having the model write any documentation at all.
It also tends to talk very circuitously and referencing things implicitly instead of just being direct and clear.
Certainly it's more capable in debugging, but I much prefer Astra over Sol now.
Of course, Astra has its own problems of code golfing everything. But at least it's very concise and direct
if you go on short form social media, Opus's ability to do agentic work in things like blender, css animation, etc. have spawned hundreds if not thousands of channels of people pumping out content with it. That and all the video game shenanigans (and now Anthropic is even modifying its system prompt to limit some of these things!)
There's definitely unique things that so far only Opus is capable of doing.
Sol 6.1 is definitely not comparable to Opus 5.5 for me.
Opus 5.5 actually decreased my Anthropic spend, because I stopped using Fable once it came out.
I ping-pong between different agents all day long and don't necessarily agree that Anthropic is clearly in the lead on coding at this point.
Opus 5.5 is amazing and has also dropped our bill dramatically. Our engineers are spending 1/3 as much as they were before and seem actually happier with the results.
Has anyone else seen this?
Anthropic can remove all access to their LLMs and start replicating entire software companies. Anyone relying on open weight models is at risk of a security breach from even slightly better models. It is ironic when people think Anthropic is the one without moat while they give all their data away.
There's a stock market truism that anyone that has real alpha doesn't sell it or give it away. I feel like there's an analogue here, that if it was just that easy for them to pack up their toys and go conquer the world, I can't imagine they would have been in the business of selling tokens at this point
IMO if llms keep improving, there's just not going to be a business in most of software anymore. It wouldn't matter if Anthropic tried to replicate all the saas companies because the category is soon to be dead. Why would you pay for someone else's software when you can prompt your own to your tastes? Every bit of software would just have one user
> Anthropic can remove all access to their LLMs and start replicating entire software companies.
"I have a girlfriend for real she just goes to a different school"
Yeah there was a vibe shift towards OpenAI being much more competitive for daily use and then Opus 5.5 hit. Will be curious to look back and see if that shows in data. Of course, could change tomorrow.
I expect growth to go down. Opus 5.5 is not only better, but a lot cheaper. My company was spending quite a bit on overages with Fable 5. Now we are no longer doing that.
Opus 5.5 is also more cost efficient than Opus 5 - costing 50% lower. Anthropic margins and/or people's usage has to be higher to continue growth.
Opus 5.5 may be occasionally better than other models on code (I find this debatable though, because Astra is very good and will often do better imo), but I would wager there's way more money to be made on things like image and multimodal tasks or at least nearly as much, and frankly even Gemini 3.8 flash can spank Anthropics best models on visual tasks and be absurdly cheaper; that also ignores the increasing aggressive classifiers Dario seems to like which block a lot of stuff and I've seen drive a number of zai subscriptions because blue teaming is increasingly very important.
Astra is not the OpenAI analogue to Opus 5.5. Sol 6.1 is.
This is likely true, but doesn’t really tell us much about future revenue growth. Now they need to compete on price, I think their revenue growth may start to look much more like typical Jevons paradox examples. I.e as their usage grows exponentially their revenue grows linearly rather than exponentially as it was before.
I’m wondering whether Opus 5.5 would have caused growth in an absolute sense or just growth relative to competitors.
Anthropic’s market share is pretty high and there is not enough total currently addressed market for growth at the expense of a competitor to make a big long-term difference.
Well what I do know as someone who has a bird’s eye view into many enterprises is that many of them are starting to realize they’re bleeding money to Anthropic and are telling their teams to be more efficient with tokens.
I've been throwing opus 5.5 at problems I got stuck on with sonnet in the past and it has breezed through them all. It really reinforced how far the models have come.
Wouldn’t previous Opus models be able to solve them too? In my experience Sonnet is often lackluster.
> I'm sure there are a lot of people rooting for their IPO to fail but they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
Then why all the rush to IPO then?
The strongest chinese models from DeepSeek, Kimi and z.ai are already good enough to run for hours on most coding tasks cheaply.
For capital to win they need to get their money out of Anthropic before the bubble deflates. This means dumping the risk on retail. There is only a finite amount of money available to buy AI stock, and the SpaceX IPO sucked a lot up, so it's a fight between Anthropic and OpenAI who is going to get what's left. The loser is left without a chair to stand on when the music stops.
Robinhood taking from the poor tired masses to give to the rich. Beautiful.
Except the IPO declared numbers on revenue end in June, so whether or not Opus 5.5 improved revenue or not, it's opaque to potential investors.
I reduced my Opus 5.5 usage and increased my Haiku 5.5 usage.
I wonder if more people using Haiku coming from more expensive models may have an impact in those metrics.
I have zero sense of this myself, am just curious, and not confident anyone really knows, but: I get that its a big deal for software developers when each new model comes out, because its perceived (correctly or not) as the next big boost for their productivity, but what about everyone else who is not a coder? Is there big reactions there to newer models? I don't know a single person who is not a engineer that would be able to tell me if "Astra" or "Opus" are Anthropic things or an OpenAI things respectively, much less when and what the big new version is for any of them. Does this play out for others? Or am I just in my own bubble for some reason?
What does average, non-computer Joe or Jane think when the next big model comes out? "My emails are going to be so much better now, I better stay subscribed"? "I can't wait to see how well it answers my questions"?
One more anecdote: People at our company are very excited for integration of existing industry specific tools and processes.
Some of them do require frontier models but many can do with localish models. Some of these processes (in our industry) are also not really limited by software at all but by the speed heavy machinery can be operated and/or manufactured and/or transported.
I'm certain LLMs are going to continue rapid growth even there but it may be an ideal field for have lesser models work for longer if it saves money in the end.
Of course who knows in the end. I'm certain there is a high chance of this being entirely wrong
I am a software developer and I don't see a huge boost since many months.
Actually I see very varying results in my work.
Sometimes I can let an AI refactor an entire codebase and then (in the same codebase) it generates absolute shitcode.
This piece was published in mid-October based on rumored numbers.
Anthropic released their Q3 numbers today which show:
- $16.96B in revenue Q3
- Up 47% quarter over quarter (Q2 was $11.6B)
- Up >1000% year over year
I wouldn't say their growth "flattened"? Their price drops in September dramatically increased Claude token consumption per OpenRouter
Sources: https://www.anthropic.com/news/s1-sec; https://openrouter.ai/anthropic
Anthropic S1 had numbers up till Q2. Many people speculated that Q3 numbers were lower because many companies had moved away from "token as measure of productivity" or tokenmaxxing. Anthropic Q3 numbers proved that theory.
As the article notes the growth is still huge. But it has slowed down relative to what they were growing.
> $16.96B in revenue Q3
The article used $16.8B. Not a big difference.
Why do we keep talking about revenue and not profit? This feels like when Michael Scott opened a paper company and the accountant had to tell him-
“At these prices, the more paper you sell, the less money you'll make."
https://isaiprofitable.com/
because revenue growth is much more important at this stage, and investors are underwriting based on the first and second derivative
We talk about revenue far later in a company’s lifecycle than we did before because high-tech, high-capex companies take a long time to become profitable.
Amazon is now the posterchild for this, but it happens in other industries like pharma a lot too. And in today’s stock market, where a fundamentally weak company like Tesla can trade at eye-watering multiples because of speculation, why worry about profitability at all? (Partly joking but that’s how people think.)
When and why we start expecting Anthropic to be profitable are great questions though.
Amazon has the world on lockdown though as the best place to buy stuff and get almost anything to my house in a day or so. I can use any AI model out there at ever decreasing razor thin margins and there is no moat or lock in at all. Nor do most people even need AI. I’ve certainly never and would never pay for AI anything. Anthropic is basically in the paper business, AI is a commodity at best, but maybe worse, because people actually need paper.
Amazon was unprofitable and not at all dominant for many years before reaching its current status. Books and groceries are also a race to the bottom of margins.
Up 47% quarter over quarter when the previous quarter was up 145% quarter over quarter and close to that the quarter before. "Growth flattened" is generous.
Everyone is really tightening usage now because of all the slopmaxers who burned millions of dollars of tokens and achieved nothing.
People tend to do wonderfully at whatever metric they're judged by. I don't think the "tokens consumed" metric was a very good one.
From what I could tell, it probably even distracted from the metrics that traditionally matter. I seem to remember tokenmaxxing dashboards being built, and other "shovels."
Completely anecdotal but I did find the better engineers tended to take to the AI tools quickest. At the beginning, tokens was probably a good way to pick them out. “Good” being entirely subjective around a product focus/mindset. If you ever find a good metric like that, it is critically important to never tell anyone, else it will quickly become not good.
> I don't think the "tokens consumed" metric was a very good one.
When mid or high level tech leadership pushes from top-down new things, that look dumb to the engineers, the leadership is either stupid or there are hidden incentives that regular people don't see. In my experience mid level usually matches with the former, and top level with the latter.
Interesting angle to blame the slopmaxers when the AI companies have been pushing maximum usage for a while now.
I assume they’d hoped that the value of the spend would manifest itself and they could raise prices. The opposite happened and companies are instead cutting back.
It sounds to me like they chose a poor marketing and pricing strategy for a product that is very difficult to make cost effective.
Everyone trying to vibecode their next million dollar idea only discover the web is dead. Nobody is looking for anything because that would require leaving the safety of the same handful of sites everyone else is on.
Because not only is there no guarantee that the peddled slop will be online or maintained or supported for any amount of time after its debut, but it's safer and more effective to just reap the slop as context into something slightly less sloppy.
Behold, the beauty of the modern internet.
What did they think was going to happen? Lots of people predicted exactly this progression of events.
For real. I'm so tired of all this, I hope the crash is swift and we can put it all behind us.
It may have something to do with the summer intern season ending. I know one of the interns hired by my employer used up the allotted AI spend by doing “fun” things unrelated to the internship project, like making Claude write an entire SQL-to-Pandas transpiler in 10,000 lines of Python. I can’t blame them for using it extensively when given access to a generous limit on token usage.
When was the last time anyone used the term "tokenmaxxing" right now?
Vibe-coders don't want to risk getting their accounts suspended if you use their models in a certain way.We are seeing the result of those subsidizations and discounts that Anthropic made this year and how much of this "tokenmaxxing" hype was used to make their revenue numbers look "good".
With the arrival of more cheaper models which are good enough, it doesn't make much sense to use Anthropic's models.
So don't be easily fooled by their revenue projections. Rather question if it is sustainable for them. (Spoiler, it isn't)
[0] https://news.ycombinator.com/item?id=49330214
Lmao a colleague who works elsewhere said some goon at work created an AI leaderboard. He spun up research agents. Top of board achieved. Granted the guys a very solid developer, but he did it to prove how dumb this is.
At the beginning of coding agents really picking up adoption (I guess 12-18 months ago now or so) I remember this sort of thing being so common that some startups were trying to sell a turnkey solution for "leaderboards".
I'm not aware of any of those products really sticking around.
We can finally dispel the myth that management is competent when they, almost as one across industries, tied continued employment to setting fire to company money
AI put a massive spotlight on all the required but unspoken decision making and coordination that must occur with stakeholders outside their own teams during the development process.
Dysfunctional middle management has always traditionally neglected it or actively undermined it, blaming others when things go wrong and taking credit when things go right.
Naive developers are always more than happy to do what they think is "picking up the slack" not realizing it's a trap.
Perhaps most organizations are just now more wise to the fact that using AI to "accelerate the process" is like arming the monkeys with machine guns?
Slop is a side effect for anyone using it at scale it seems.
It's a combination of natural language being an awful programming language and frontier ai companies optimizing for longest time horizon at 50% success. I mean what did they expect? Length of time to not check in and still bumble around to turn a single prompt into a literal half-assed solution.
I now think of software code as DNA.
Your machinery was not written elegantly. It was sloppily assembled via evolutionary pressures.
A gene gets duplicated, mutates, and suddenly you have new functionality and new fitness criteria.
Try something bad and you die without reproduction.
Our code isn't clean. There's redundancy and extra packing and slop all over the human body.
Embrace it. Stochastically hill climb.
Code is DNA. Evolve it, don't intelligently design it. Fit the curves of your environment quickly and use tests to cull bad behavior and the slopes of your domain.
I know this is tech sector funny money and we do just disregard profit and judge companies on revenue alone, but the 5.5 models that just released are quite a bit cheaper than the 5.0 tier.
I wouldn't really be surprised if better model efficiency was part of the reason for a revenue stall, but I would also expect their costs to have dropped a fair bit too (although I have no idea whether that corresponds to an actual cost reduction or if it's all tied up in hardware and land that they already bought and will need to hit more utilisation of now.)
It's fun to note that nobody mentions AGI in these threads anymore.
I think most people who care or are tracking it would say we've hit it, and those who disagree just have their goalposts a lot more clearly defined and are quietly waiting.
The very first goal post for AGI must be that it can act, learn, and grow independently. They must be self healing, self correcting, and anti fragile. That's definitely what the G in AGI is about. But all models fall into decoherence if you let them work on their own too long.
If you left an AI alone for a long time and when you came back they had built up a working society and learned new useful things, then that's the proof it's AGI. From this definition you only know it's AGI after you let it loose.
I don't understand how anyone can call these AGI. I've always had the view that AGI == Independent human-level intelligence.
At some point someone moved the goalposts for this to include LLMs that emphatically are not that. I think it was Altman when they came out with Astra?
i don't think it will ever be comparable
they aren't as intelligent as a human, but they can throw shit at the wall faster, and get something that sticks faster. they are superhuman in areas that require trial and error, like hacking.
everyone stopped talking about the Turing test once ChatGPT came out and broke right through it while still being very very dumb. AGI is just too blurry to define.
I think it will probably happen someday, but not anytime soon (and probably not in my lifetime)
Because we already achieved it, haven't you saw the post by Nvidia's CEO Jensen Huang?
Nvidia's Jensen Huang says "We've achieved AGI."
It's because all the arguments have already been said, and now neither side believes they can convince the other.
If the hypothesis is that "everyone is tightening spend on usage", then anthropic is boned, if that holds true. Look at the price to performance for any benchmark and compare OAI and Anthropic, and this would be a pretty clearly terrible reality for Anthropic. OpenAI absolutely blows Anthropic out of the water when it comes to cost, if you pay per token.
I am guessing this just means OpenAI is running at a bigger loss than Anthropic, not that that improves the situation.
Or it secured more compute early at a lower price.
Also model/token efficiency.
List of possibilities is big
The real question is how many people haven't figured out how to use AI to make their job easier/more productive.
Especially with this latest crop of models, they seem to be pretty damn competent in classic engineering rather than just software engineering.
> The real question is how many people haven't figured out how to use AI to make their job easier/more productive.
I think the answer, to a reasonable approximation, is “everyone”. Outside software engineering most people aren’t using it for anything useful, and I keep encountering technical people including programmers who don’t use it either, which never fails to surprise me.
One of my projects is a transportation management system (TMS) and of course I have added a lot of features to it including a very capable chat-based UI that can automate a lot of tedious data entry and other tasks for dispatchers and accounting and so on. It gets used but really not very much. I’m starting to realize that I’ve probably taken updates as far as they can go in terms of the ability of people to absorb change. Most people just carry on doing things the way they always have.
We had one person on the admin team who started using Cowork to automate all kinds of stuff. She loved it. Then she left. It hasn’t caught on with anyone else.
This is not to say that things won’t change but I think it will take a while.
my brother uses it to do ocr on his books but, lately, claude has been refusing because of copyright or whatever so i've had to do it for him (using ai, yes, but not anthropic's)
I know firefighters and machinists heavily using it, the only places I see slower adoption is purely because of bad leadership- exclusivity contracts with certain providers or just plain fear.
I think this dynamic is underappreciated. Another thing I see in working with different companies (consulting) and discussing with colleagues at different jobs is that we as humans tend to do a pretty bad job of modeling business outcomes and the dynamics that govern them. There are lots of outcomes where you can identify one contributing role, increase the productivity for that role by a significant factor, and still ultimately not move the needle much, because the constraints span multiple functions.
As an example outside of software engineering, a lot of people in marketing roles have found agents to be incredibly useful for particular tasks. I've talked to multiple people who wouldn't standardly be classified as "technical hires" who are using Claude to manage things like digital advertising. It's valuable here because managing campaigns on a platform like Google or LinkedIn involves a ton of granular tweaking and analysis. Every performance marketer makes tradeoffs around what they're realistically able to spend time doing, and what makes the biggest impact. So if you have 250 campaigns that you manage solo, you likely don't have the time to really manually optimize the bidding and budget for each campaign every couple of days based on new data. You can only try so many new angles, because creating new media is time consuming. You don't have time to construct custom experiences on-page for every particular campaign. There's a whole industry of tools that turn a tidy profit automating parts of this for this reason. And agents are really good at it, because a lot of it is just applying the same analysis over and over again.
But for a lot of startups, increasing traffic from ads 20% while decreasing spend is a good outcome, but it doesn't actually turn into revenue. The product itself might just be limited, and already being discovered by the narrow segment of potential users who find it valuable enough to pay. Or the sales process might be kind of busted, meaning the team will really only close a certain volume of new deals in a given month, no matter how much inbound they have. Or some combination of all of the above.
If it's just one marketer using an agent to optimize their advertising, there's not a lot they can do about the other constraints. And so what you end up with is one employee who feels really productive and sings the praises of Claude, but not a meaningful lift in the higher level outcomes you care about, and the analysis gets messy. In an ideal world, everyone realizes where the other bottlenecks are, but that's a genuinely hard thing to do and people have complicating incentives.
I don't think this is a particularly new problem. It feels identical to any other situation where a department asks for a tool, loves it, and it doesn't really make a difference to the bottom line or make anyone else's life easier in noticeable ways.
> very capable chat-based UI that can automate a lot of tedious data entry
Because rewieving chat bots data entry is even more tedious then doing it yourself. And as long as they have to review it, it does not even matter that it makes only few mistakes.
I see it with JIRA and Rovo. It just sux using it, despite it being reasonably competent at creating tables and what not. If I could do markdown as used to be normal, it would be appreciated a lot more.
A lot of people have jobs that aren't limited by how much "stuff" they generate.
We are capped by testing and PR review. The junior devs can spam out a million PRs a day but we keep having production incidents from not carefully scrutinising the changes. So the speed of code generation is limited by how fast we can actually accept those diffs.
Production incidents from not carefully scrutinizing changes is a sign that someone more senior should be putting infrastructure in place to minimize the impact of those incidents.
Feature flags, staggered rollouts, quick rollbacks, smoke testing, etc.
Sure, if you were hand managing every rollout up until now, and relied primarily on unit tests, you’re in for a world of hurt when you give a junior engineer Claude code and let em rip.
But if you have more advanced systems to measure and maintain overall service health, and codify your business logic invariants, you can really start flying these days as long as you’re all aligned on what to build and the potential impact.
I agree, with a big tech deployment process this would be less impactful. But getting from a company used to manual testing and weekly deployments to a fully automated CD pipeline is more work than a single Claude prompt.
>someone more senior should be putting infrastructure in place
High chance said senior has been laid off because "A junior with $200 claude plan" is cheaper.
Don’t forget requirements gathering and design. There’s only so fast the business side can feed requests into us.
The business side will step up their token output.
Haven’t you heard, Anthropic uses Claude for all of its PR reviews and has no issues at all!
Unfortunately I work with users who get upset when the app goes down or features break.
Unironically you should be using claude for monitoring and within CI. "PRs causing bottleneck" is an anti pattern during AI. It is as easy to verify as it is to generate.
That was always the issue... Rewrites were done before LLMs and the results were the same... There was never a point in time where the bottleneck was production of code, it was always review, testing, QA, fixing bugs and edge cases, and that is when only focusing on the technical part of delivery, and leaving out communication and synchronisation within the company and between company and clients...
That strongly depends on context. In larger organisations, this is definitely the case. For smaller startups and teams then the overhead is much smaller and the actual core programming work is much more of a bottleneck.
Sometime in the last 15 years we got these huge backlog wishlists of one line thoughts. People were convinced because software devs couldn't keep up translating one line thoughts into working software that the code must be the bottleneck
They never realized how much effort went into turning those one like tickets into code was, and the code was the least part of the effort
How so? Or is this just vibes based analysis? I dont really know what you mean by classical engineering, but sure if something can be done on a computer it can infinite monkey theorem its way to a solution. This isnt exactly useful its just an enablement of laziness.
Idk how many people need to hear this, but not everything 'is computa',no matter what Trump says.
Much of the worlds problems are not verifiable with a compiler that can be trained or brute forced against.
Llms are not as useful as ya'll think they are, not everything exists in a web browser, which is why Anthropic et al have saturated their TAM and are now frantically hiring sales people to convince managers in other verticles.
Turns out most people dont need or want AI.
Right, this is what astonishes me about the constant rhetoric/hyperbole about how this is going to impact "all the jobs" and make everyone wildly more productive than ever before. It's a fabulous tool in a limited number of areas, and it's otherwise only handy here and there. But just because it's a fabulous tool for the coding that is so prevalent among HN users does not mean that it's a fabulous tool for everything.
A hammer is fantastic for driving nails, but it's not really all that good at washing a dog.
Name a single job that can't be improved with better communication, planning, or insight.
This is conflating the benefits found within those areas with the ability of AI to actually deliver those things effectively across every single profession. Improving something by 2% vs. revolutionizing an entire industry are wholly different propositions. Maslow's hammer would like a word.
I'm not sure I'd rank LLMs very highly at any of those on their own. In the hands of many users they can even help wreck havoc with those processes.
Fixing the shitter
LLM make communication worst.
Mechanic, welder, builder, farmer, logger, millwright, trucker, heavy equipment operator... I could go on. Literally the entire economy in the area I live is completely impenetrable to tech, let alone AI. None of it is worth a shit to any business here.
These are all jobs the silly con valley idiots claim continuously to be automating away, but it's all just fuckin lies.
The extent of tech's penetration into this economy is internet ads on facebook and streaming services. Nothing to do with actual work.
I use classic engineering to refer to the (engineering) disciplines where mother nature messes your day up rather than obscene levels of man made (or machine made) pure complexity ruining your day.
It's easy to understand how LLMs got so good at coding, they can rapidly iterate, self test, and it all happens in their digital universe.
These models now starting to spit grey beard knowledge of precision analog electronics, lightweight fastener design, and chemical surfactant systems is something very very recent. A few years ago I think we were pretty confident that the tech would hit a wall at the digital divide. Well, I can tell you now that it did not. It seems it kinda blasted right past it.
Vibes based analysis is very high volume right now due to upcoming IPOs.
And how much of that is astroturfing, one should ask.
[flagged]
On HN, can you please not be snarky? We're trying for something a bit different than internet mean here.
The idea is: if you have a substantive point, make it thoughtfully; if not, please don't comment until you do.
https://news.ycombinator.com/newsguidelines.html
AI growth is unlikely to be linear. We have seen a lot of hype in the last 12-18 month that give us ideas like "tokenmaxxing". But this is the same cycle that we have been before: hype -> disillusionment -> someone figures out a better way to use it -> next hype phase starts.
As it stands, even the people closest to this technology have barely figured out how to use it effectively.
Sometimes it's just Hype -> disillusionment, and then no one ever figures out a better way and it's dead technology.
Not saying that applies to AI, but it certainly applies sometimes.
I can hear the landfill full of those 3D glasses for home TVs shout in agreement
It's really hard to bring everything there on my Segway, which is of course the only type of vehicle on the roads anymore.
Plus all the blockchains.
AI has historically had a habit of disappearing for ten years and coming back once memories fade and the new story sounds enough different from the old one.
I don’t think this era is going to end with us all unemployed, but I do plan to be retired before the next cycle.
As an investor, what's the moat here? Being 6 months ahead of China's open source/weight models?
What is the moat of any tech company if they are using LLMs to build their products which can also be reverse engineered by an LLM? Hardware?
For big ones it’s often network effects and/or friction. Those can be pretty good moats even with LLMs. These models are also nowhere near the point where they can actually reverse engineer complex products.
Network effects have subsequently been; shown to be logarithmic. You’re not getting 100x the benefit from 10x the users. You’re lucky to be getting 12x the benefit and you also have logarithmic increases in costs because there’s no such thing as economies of scale in software. You pay for every order of magnitude in hardware or software development costs.
It was six months last year, it’s on the order of weeks now. Presumably they have unreleased models that are a step ahead still. The moat is capacity and how much of an ecosystem the can build around the llms; eg anthropic figma tool.
Oh how badly I wish this was true.
How much easier my life would be if Chinese open models were even just 1 year behind outside of X posts and one-shots.
My easy litmus test is would the year old model (Opus 4.5) being open sourced today would eat the traffic of every single Chinese model... and the answer is pretty much yes.
Every Chinese model would throw away their work just to post-train or distill Opus 4.5 and get a more reliable model.
Having compute to serve the demand.
I always bought on who I thought was best at attracting the best talent in the space, and for how long they could remain dominant in that.
decent software stack that works out of the box.
You can go very far in life with a thing that isn't futzy
OpenAI's Astra released on Sept 3 and Opus 5.5 released on Sept 22.
Most likely the combined revenue growth continued just as fast, but OpenAI took a lot of market share with a model that's ~as good.
If true, then the current market sentiment is too pessimistic because demand for AI (and chips) has continued growing.
I think the next cycle of growth will be from increased subscription pricing? Thats what the streaming companies did - start low, keep increasing forever. OpenAI already merging codex / chat and I think most people will have to buy that min subscription or even the higher tier. I am already at the $100/month
I am not saying its a must, but those who are reaping the benefits in someway will end up spending more I guess?
I think they'll take a page from OpenAI's playbook, namely introducing a higher priced tier, and changing the limits on the $100 and $200/mo plans.
I feel like a lot of this could be explained by how good Opus 5.5 is compared to Fable. People don’t need Fable for most tasks.
Opus 5.5 is basically not included in this data.
They'll be competing on price now with openAI, google, and Amazon eventually.
AI is just a compute commodity with a value add.
Anthropic and OpenAI own no compute. Amazon has no models.
So that leaves...
They can't keep getting away with it.
At this point only they have the skill to f' it up from here.
Amazon hosts models. And?
Amazon has to pay ~70% revenue share on models they host.
Not the open weight ones
Are any of the AI companies selling their services at a profit or are we all being subsidized by investor money that is about to need to be paid back?
This is naive. AI is one of those few things that does have moat. Its shockingly hard to hire people and create a good company culture that does frontier training. There's only like 3-4 labs in the world and the rest are distilling.
Even in next 3-4 years, the company that has the best ecosystem of researchers, culture, shared knowledge and processes will be able to churn out models in a way that can't be replicated.
i think its a lot deeper than that, compared to deepseek-4.1-flash, glm-5.3-flash and the recent minimax-m3.1-flash [0], the comparison in price (15x cheaper inference cost) and the productivity gains are very similar if not more profound depending your prompt topic (cyber, biochem, etc)
[0] https://x.com/jumperz/status/2112226308868436154
I don't see the issue. If the market is saturated by them, the slowdown seems to indicate to me that it has reached (at least transiently) its intended audience. If the end goal is that every person on earth has a claude subscription (or whatever your choice is) then yeah that is worrying, but I don't deem that realistic.
They need a lot more growth just to afford the compute they've already agreed to buy. Not to mention to justify their rumored record-high IPO price.
Why do you say this? If they get to even 100B RRR they will pay off all of their commitments in <10 years with money to spare. Does that seem so impossible to you?
No one with a brain does, but Anthropic’s valuation is banking on massive sustained growth to pay their bills.
It's also very damaging to under-value, I think OpenAI got hit hard with lack of compute and is still struggling.
Flattened might be an exaggeration but they lost many a customer with boneheaded decisions like blocking subscription access to OpenCode. I don’t think they’re past the point in their growth curve where tinkerers don’t matter anymore. Again, definitely not flattened but not as parabolic as they could be.
Zitron has been saying this for weeks.
EDIT: okay, not months, weeks. Feels like forever.
[flagged]
Please don't post swipes about people on HN, no matter who it is or how right you are or how right you feel you are. https://news.ycombinator.com/newsguidelines.html
Even a broken clock is right twice a day
Years, no? Meaning the sky is falling rhetoric was wrong for a long time. He will get lucky eventually.
He says a lot of things, but one specific thing I remember from weeks ago is that he said Anthropic growth slowed on Q3, and that's why they decided to push for the IPO (and use only Q2 numbers).
It reminds me of the Apple TV rumors. For ~4 years the new AppleTV was just around about to be announced for <reasons> and you should wait to buy one. Then it's finally announced and people are all "yeah, but a few weeks before the AppleTV was launched they said it was going to have more RAM!".
Sure, and they said similar kinds of things for 4 years. Why act like the broken clock is suddenly worth noting just because the time was finally right this moment?
my simplified take is the headwind is on two ends - smaller models being sufficient enough & open weight models then second utility of 'a.i' is maximal when deployed with humans i.e engineers applying a.i on top of current applications putting an agent layer on top.
the second part is not sexy - but we see all a.i labs having fde arms e.g ode, microsoft frontier company etc but then again that's the part you can't hype.
This is fine. Not only are we happy with growth now, but we want growth of growth?
The conclusion that this is a bubble seems arbitrary and disconnected from the several true facts outlined in the article. Anthropic's growth _is_ slowing, albeit still strong at ~47%. They do have enormous compute obligations, and it will require a lot of growth to meet them. However, they also increased revenue per dollar spent on compute from $0.40 to $2.08 in five quarters and hold $131.7bn in cash.
Keep in mind this growth has mostly been sustained by low-level business automations, coding, and chatbots. With the adoption of products like personal agents, e.g., Meta's Muse, I am pretty optimistic there will be demand and growth around this level will be sustained for a few quarters. Doubtful we will ever see another ~150% growth quarter.
> revenue per dollar spent on compute from $0.40 to $2.08
So that’s ignoring labor and perhaps facilities? That’s like saying we’ve hit net positive fusion by only considering the power produced by the laser and ignoring the 80% power losses in the feed-in.
I am actively pushing my workplace to scale back our AI usage (claude cowork)... the amount of AI slop being pushed between teams is off the charts and ends up creating a mountain of work for reviewers. Hurts productivity rather than improves it.
And what is between June and September?
- Would you like to get paid a million dollars right now, or get paid a penny today that doubles everyday for 30 days.
- hmmm, why 30 days? Make it 60 days and I'll take the penny.
The effort needed to double everyday, or month is realistic at first. But then you can't just keep doubling without swallowing the universe along the way.
The answer is ~5 million in case you’re curious how much the penny is worth after 30 days of doubling.
~5 quadrillion after 60
You could in theory be responsible for the destruction of the entire solar system in order to pay you. Sit on your throne atop a mountain of skulls and weep.
The mountain here is paperclips
For anyone that hasn't seen the game https://www.decisionproblem.com/paperclips/
You can tell how desperate an argument is by how many derivatives they are using.
Growth being the first derivative… so… not particularly desperate?
Their growth was positive. The second derivative (growth of their growth) was likely negative, and that's what this article is about.
Right, but a comparison is always one more implicit derivative. If you look at revenue itself, it's growing or shrinking, that's first derivative. So there's one more derivative here than usual?
I guess the interesting thing would be to look at actual company assets/value rather than revenue; revenue already in itself is a derivative. I.e., contrast their revenue against expenses. Last I heard those numbers are deep red, aren't they?
Revenue is an integral. They sum up the rate of cash in versus cash out to get revenue. Thats dollars/hour over 91 days.
Growth is first derivative. Deceleration is second derivative. Focusing on second not first I think was the point, not implying there is some nth derivative under discussion.
Winning an argument online 101: call your opponent desperate.
I have no issue taking arbitrary derivatives... provided you're not taking derivatives of very noisy data.
when politicians and bankers talk about "reducing inflation" they use the same derivative
New bubble pops theory found: the 15th derivative is negative guys! Ed Zitron's 10000 word slop essay to follow
I thought growth has slowed due to crackdown on distillation
They said distillation was 5-7% of revenue. The slowdown is much larger than that.
Insane to invest in that company on IPO
Insane
It's because everyone is talking about how the bubble will burst! When people talk about it a lot, they tend to believe it. Everyone just needed to give their computers to Sam so that no one could talk about the bubble, and then everything would be fine.
Main discussion:
Anthropic reports $17B in Q3 revenue, commitments of $895B in revised S-1
https://news.ycombinator.com/item?id=50337439
DRAM ...
If you ever wonder why Anthropic is going all out to try and fearmonger their way to an open weights ban, this graph will provide all the answer you need.
I think all numbers coming out of Anthropic are utterly meaningless until we can answer the question of "just how subsidized are those subscriptions?"
Their financial statements suggest the answer, even when you take into account the cost of training the model, is now 'not at all'. They are profitable, the main questions for their valuation are 1) can they keep their margins, 2) how big can they grow?, and 3) can this growth match what they have committed to?
Question one is how do their investors get paid back. I found Series E through Series H totaling $111.5 billion in investments. Since series E was $3.5 I’m assuming all the previous rounds added up to less than that, so let’s say 115 billion at the outside.
That debt will transfer to public shareholders at the IPO but also will double or triple in the process, as the investors expect to be paid very, very well. So the current shareholders are looking to extract something over $400 billion from the IPO.
With the current revenue they might pull off that P/E. But they’ll still be extracting some of that value from the customer base, because shareholders will demand it.
Not at all, the subscriptions have an adjusted gross margin of approximately 85%. Very profitable! The training is a different story though. How long will the subs last if they stop training? Also, why on earth would they put on the breaks and try to be profitable when they have plenty of funds and are in the growth stage?
The subscriptions are a key part of the engine to ensure The Machine can keep improving itself by harvesting all that data and feeding it back in. Many of the enterprise contracts forbid the lab from training on their data, so the subscriptions are a form of gold.
The contracts might, but do they pink promise?
The company was founded under the principle they will violate laws to acquire training material
This is another story, and I'm inclined to agree with you. They are desperate for more data.
The subscriptions do not require training on your data.
The exact meaning of the "don't train on my data" checkbox is pretty hard to pin down.
> all numbers coming out of Anthropic are utterly meaningless until we can answer the question of "just how subsidized are those subscriptions?"
Subscriptions are a growing minority of revenue.
their gross margins point at this not being an issue.
Ah the beauty of circular financing. By owning the company who pays me to provide a service, I stand to gain massively by giving them a discount, making them look profitable, and fooling investors in to providing my exit liquidity.
Which of Anthropic’s revenues are circularly financed?
Is ChatGPT down?
If you can’t answer this, perhaps.
The criticism has been about folks downstream of Anthropic being circularly financed. I haven’t seen credibly claims of Anthropic’s revenue being circular, but if you have a source, by all means.
Doesn't their S-1 render this irrelevant? Their financials are clearly much better than we thought
Anyway the growth this year is incredulous!
Growth is conscious now? or what do you mean?
With AI who can tell.
Huh? They provide a service that does something that did not work before and now does and was quickly adopted. They do not have a moat.
I am just imagining Anthropic (and OpenAI) employees reading these comments and laughing at the rank ignorance of the commentariat.
I agree, share you system prompt here so I can mark this as a success.
I've been saying it for a while now but if someone sold me a box with model perf from ... I guess November 2025 then that's decent enough for my usage. Tool interaction is the magic, I don't need much if any more innate intelligence. Things are fast enough.
I wonder if Anthropic will "see the light" and stop trying to burn a bunch of money and sell a thing they already have (which is decently useful). Very ready to be proven wrong in any direction
wow really? Nov 2025 I was still fighting blatant hallucinations in helm charts. Now I'm two prompts away from a new build published on the app store and 5 video reels in the queue complete with captions and everything
The thing is I'm not spinning up a new app every day, I'm just working on a large existing system with enough constraints. From that perspective tools + the general rigor I already try to apply means that I can get some value out of LLMs without waiting for it to do something "good".
"I can one-shot a new thing" is interesting but definitionally a thing is new only once. Maybe if I worked in a space where we had way more throwaway things to work on.
At the end of the day I want to understand the output anyways, so it one shotting a thing vs me going back and forth with it in smaller increments is a mixed bag anyways.
I think one thing that could explain it is that some of the big tech companies might have been big customers of theirs, and now they have alternatives that, according to recent news, their employees are being pushed to use instead. Some of those reports mentioned internal usage getting cut in half. If this is true, it doesn't necessarily mean their growth has stalled.
Discussed here a few weeks ago: https://news.ycombinator.com/item?id=49997161
Employees were and are pushing for more model and harness diversity too I promise.
But perhaps for different reasons than Microsoft, Meta, and maybe Adobe leadership are now. See, if they are all in on Claude how do they sue Anthropic for assisting in copyright violations and IP theft? The services would be cut immediately.
Well when they don't need them anymore..
This is why Anthropic is heaping on more and more model restrictions running up to IPO. I believe it's a smokescreen for the IP risk they outlined in their IPO filing..
The mental gymanaatics here is crazy
1) customer concentration risk 2) when those customers use a diverse set of models - so too will the rest - they’re not idiots who want to pay more than they should
Aka % growth rate decreasing sharply and eventually flattening out. Within 2-3 quarters.
Understood?
This is why Musk rushed consolidating xAI with Spacex and rushed to an IPO. And, why it was so important to him that Nasdaq and Russell rewrite their rules so he could dip into everyone's retirement accounts as quickly as possible. I don't think he's particularly bright, but he has an incredible cunning about how markets behave and how to manipulate those markets to his ends.
The writing was already on the wall about a lot of things WRT the AI market: 1. there's a limit to how much money will be thrown at it before investors expect a return, and that limit is fast approaching, 2. it's turning out "intelligence" is going to be a commodity and it's not much of a moat, 3. cornering the market on memory and GPUs (even if they sit in a warehouse for months or years before going online, as long as it drives up the price for competitors) only works if your competitors don't figure out how to do it more efficiently and on alternative platforms.
Spacex also soaked up a lot of the retail appetite for AI, even though it's an also-ran that barely qualifies, in terms of enterprise and government usage (and for good reasons, both technical and otherwise). Every other AI IPO is going to be facing a more pessimistic market, especially if we continue to be in a tough economic environment.
In short: They waited too long, now they're in a world where they've burned so much money people expect them to deliver something that no one in the AI industry probably can deliver: a monopoly-like hold on a huge percentage of the market, and to invent a bunch of new markets that aren't currently buying AI. They have to have a dominant position for this to work because they have to be able to set their price higher than is standard now, without everyone fleeing to other cheaper providers. And, when the market for memory and GPUs finally gets back to normal and supply catches up with demand, I'd wager a lot of companies, and even individuals, are going to get real serious about open models and self-hosting (unless prices remain competitive for the hosted models).
> 3. cornering the market on memory and GPUs (even if they sit in a warehouse for months or years before going online, as long as it drives up the price for competitors) only works if your competitors don't figure out how to do it more efficiently and on alternative platforms.
My understanding is that the GPU market is largely sold out for the next few years?
What exactly is the threat here? Like, which specific competitors, what efficiency improvements, what alternate platforms?
I didn't think I needed to spell it out for the HN crowd, since folks talk about the Chinese models every day. But, OK...
At least three Chinese companies have bragged about their latest models being trained and/or served on Chinese-made GPUs. Literally within the past 24 hours, LongCat spent a large portion of their release announcement for 2.5 bragging about how efficient their training process is and that it runs on domestic (Chinese) GPUs: "LongCat-2.5 uses long-sequence training during both mid-training and SFT. With less than 80 GB of memory per domestic AI accelerator, the training system prioritizes memory efficiency and aligns its parallelism with the supernode architecture."
https://longcat.ai/blog/longcat-2.5/
DeepSeek has repeatedly blown the curve in terms of how efficiently a model can be trained and served, Xiaomi and Z.ai aren't far behind, and have also discussed training and serving on home-grown hardware. I'm sure you can search Arxiv for all of the papers from these providers (DeepSeek, in particular, who document seemingly everything, even when they're doing something ahead of the US providers) as well as I can.