My chatgpt pro account gave me 62k of their "credits" which they try to sell you at extortionate rates when your subscription runs out. I spent two days hammering it and used about 4k of those.
This is clear manipulation to prevent me from complaining about lower limits until it's out of the news cycle, but I don't care. As soon as those are gone, if my $200 account doesn't give me the value it used to, I'm out.
The random resets are also clear manipulation in this regard as well. Rather than just give me a fixed weekly limit which I could then get a sense of and know if it's getting reduced, they throw out a ~0-3 resets a week at random, unexpected intervals so that I'll never know.
I don't think it's them exclusively.I'm the admin on the claude account at work and I have their cheap account and never hit my limit, while others are on the expensive plan and run through hundreds of dollars of credits in addition each month. I'm not taking it easy on this thing either. I had it convert several thousand rules in our firewalls to fortimanager global rules, I gave it a personal project with a long promp to build a truenas replacement based on universal blue's ucore, webzfs and config files for nfs, cifs/samba, and s3 (rustfs) all managed through my forgejo setup with forgejo runners building it and forgejo providing the repo that the machine looks at to upgrade the base and the config (it's two seperate pathways so I can push config changes like new nfs shares without having to do a system image update). All that barely moved the needle on my account. I suspect they are all a bunch of shysters opportunistically abusing us.
> I suspect they are all a bunch of shysters opportunistically abusing us.
At this point we conclusively know that.
What frustrates me is the constant unshakeable belief here that subscriptions must be running at a loss, based on zero evidence.
Anthropic have said they have 85% margin, which presumably means their normal API pricing. 85% margin is insane. No business deserves that. We can easily assume the subscription is running at a more sane margin like 5-15%.
I would pivot to a Chinese provider via open router so fast... Most months on glm or DeepSeek it would be a challenge to spend $100. But now I mostly use local models... Sounds obnoxious bouncing between 2 companies who rug pull their customers every 4 months and degrade service as a form of advertising...
Do you have beefy hardware? i dream of using a local LLM, but neither my 24GB RAM M4 Air nor my PC with a RTX3080 and 16GB of RAM seem usable (yet).
i could label them usable if "prompt it and then wait for 6 minutes for it to change a single line" is deemed usable, which i don't.
spent the last weekend in a rabbit-hole of which models to use in what kind of setup, but i think with my hardware i'm just out of luck for now.
for private matters I can use Cloud AIs, but not allowed to use it for work-related matters, which is where i could use it the most. they do provide us an isolated AI environment where we can use Claude etc. for work stuff, but heavily rated (about 20 prompts per week per Claude model).
You're super close on your PC. It seems like the sweet spot right now is 24gb vram + lots of RAM for qwen 3.8 flash next or a dgx spark like machine with buckets of unified RAM.
You could run some decent models on your PC l, but nothing that would totally replace serious work. Qwen 3.7 35b is pretty okay. Maybe you could try the new bonsai quant for 3.8 27b but I doubt it would be awesome.
i did in fact test different Qwen's! :) qwen3.5:9b and qwen2.5-coder:14b, i think.
but i struggled with speed and them being stuck in "Hmm, wait, no, that doesn't work, let me analyze again"-loops. on the M4 Air I was able to start some bigger models. definitely the first time i burned through 25% of battery in half an hour while getting first-degree burns on my lap.
maybe in 1-2 years my hardware will be enough for fast local-AI. maybe by then i can afford new hardware. (think with current trends option #2 gets less realistic each quarter).
not sure if i lost a bit of curiosity and spark. there's so many models, tools, harnesses, tweaks for harnesses, edited models and so many things. ultimately i don't care enough, it seems vapid to stay bleeding-edge informed about LLMs, if one's career does not directly depend on it. i mean, what the fuck even is a bonsai quant, this stuff makes me feel like an uninformed excel boomer even tho i absolutely don't am one :D
not a software engineer per se so i don't need 16 agents running 24/7 with openclaw. i want a local buddy who helps me write ansible/python faster, better, helps me analyze bugs, refactors small things. basically what Claude Web does, with the benefit of it reading the files itself and running locally.
maybe i just ignore the whole AI/LLM part and just buy some new hardware so i can use higher graphics settings while gaming, WAIT NO, that market was nuked by AI as well, guess i'll continue using my i5 from 2019.
> subscription gross margins are way lower than API and meaningfully reduce revenue per MW for both OpenAI and Anthropic
I don't know about everyone else, but there is no way in hell I would pay even $200 for API-priced tokens for personal use. So at least for my sample size of one, Anthropic's revenue would not be higher if they dropped their subscription plan (they would get $0 from me instead of $200 per month).
I’m not everyone else but I am SOMEBODY else and you’re absolutely right. I was paying for Claude since they first offered subscriptions, right up until this summer.
I respect Anthropic (and OpenAI to a lesser extent) but I’m not going to play these games.
Local ai can be about as capable as anything you get a subscription for. The benefit to local beyond a lot of other things is, no model degradation, and stationary costs. Actually, you get effectively free model upgrades.
For 2-4k you can have opus 4.8 at home running faster than anthropic. In 8-20 months you have broke more than even.
I think for almost anyone it's worth trying. Especially if you already have hardware.
Yes, but we are getting close to the point where the difference exists, but it is less and less significant. Tbh, Qweb27b is perfectly find for a lot of tasks; anecdotally, the difference lies less and less in quality of responses, but in speed. Local models have a hard time there for a reason ( the cheapest stuff you can get now is stuff like previously mentioned 128gb ryzen ai ).
So trade-offs are there, but do those matter to all people the same way ( are they not equal in the same way to everyone )?
I would say from experience, that yes the local models today are as good as the old frontier models. That was my original claim. I am sticking too it. It won't be long before local models match say fable 5.x as well. And people will still say "it's not as good as x". No it's not but it's as good as the thing everyone freaked out about 6 months ago. If that wasn't good enough for all y'all then then why did you all use it? Exactly.
This is the elephant in the room that OpenAI and Anthropic specifically aren't talking about -- when "good enough" is surpassed.
Because there are a large number of use cases (and large sub-portions of others) where genius-level AI isn't required.
Which means once that threshold is surpassed in people's relevant fields, available margin on that is going to collapse to commodity levels.
It's difficult to see how either pure-play AI company maintains its valuation once that happens. Their TAM is based on capturing a big chunk of all work, not just that which requires the highest intelligence.
$4k doesn’t even buy you a DGX spark these days. You can get a Ryzen AI max 128GB machine for a bit less than $4k but the models you can run on there aren’t even close to Opus 4.8.
Amazon recently moved $8B in GPUs to a SPV to hide away their expenses. Oracle warned their data center isn’t going to be on schedule. Worst of all, none of the big players are reporting enough deprecations to match up with GPU sales.
All this indicates to me that loads of GPUs were only purchased on paper or are sitting in warehouses unused. This almost certainly means a drop in orders followed by a drop in RAM prices (though that would indicate a demand drop to investors, so maybe it’s better for stock prices to keep paying to bills bigger warehouses and stuff them full of unused GPUs).
The second RAM prices normalize, local LLM becomes much cheaper. A machine that doesn’t make much sense at $10-15k suddenly becomes a lot more feasible at $3-5k. If those stored GPUs flood the market, we might see even bigger price drops.
that's a wild theory. You think there are really warehouses of GPUs that they are just sitting on until data centers are built, that will just all of a sudden flood the market?
First, the form factor of GPUs in data centers aren't the same as desktop GPUs, so you couldn't use them even if you wanted to in a normal rig.
Second, from a business perspective that doesn't make a lot of sense. It's much more likely that GPUs are going to the highest bidder/large contracts who are scooping them up to populate/upgrade data centers that are in operation because they are going to get more money per gpu on newer hardware. The "old" hardware might go to a warehouse to be auctioned to the highest bidder or go to a data center coming online, but I highly doubt they are sitting on market wrecking amounts of GPUs just waiting to flood the market. When they go bankrupt and they have to sell a datacenter or two, those datacenters being parted out as part of a bankruptcy deal I could see. Warehouses full of unused GPUs doesnt make any sense to me though.
> First, the form factor of GPUs in data centers aren't the same as desktop GPUs, so you couldn't use them even if you wanted to in a normal rig.
With eGPU PCIe 3.0 or 4.0 x4 links (or even Thunderbolt), that essentially doesn't matter for inference.
You pay the bandwidth hit on model load / unload (a few seconds), but it's irrelevant for post-loaded inference.
Anyone hosting a high wattage GPU is going to be fine hosting in an external enclosure, most of which have generous extra-spec room.
And if they don't, if a flood of cheaper DC GPUs hit the used market, you can bet Chinese manufacturers will have enclosures that fit them available the day after.
It makes sense if you consider that maybe all of these companies are colluding to forward a shared agenda. It is happening whether it makes sense or not. So there's that too.
>>The second RAM prices normalize, local LLM becomes much cheaper. A machine that doesn’t make much sense at $10-15k suddenly becomes a lot more feasible at $3-5k. If those stored GPUs flood the market, we might see even bigger price drops.
Yeah but it's unlikely to happen in coming 2-3 years (at least) and then the SOTA models are going to be a completely different animals.
I would love for this to be true, but this is nonsense. The local models that you can run for those prices are not as capable as current frontier. At $200 a month, it would take well over a year to break even, and that’s if you don't include energy prices and time spent tinkering to get and keep it working.
I exclusively pay for API tokens. This is the primary way I consume these models.
The flat monthly subscriptions seem too tempting for the providers to screw with. I prefer to paygo and to be responsible with my consumption. I also want the ability to scale substantially beyond what a typical consumer plan may offer on occasion.
My monthly usage ranges from $10-$1000. I don't have to worry about quotas or anything. If I need to use several thousand dollars worth of tokens, I can just pull out the Amex and everything works. It's constant performance all day every day. I have long since maxed out my org level with OAI, so it would be very difficult to exceed any realistic limits.
Spending 5/10 months equiv of subscription cost that have significantly lower revenue for X company (or even costs money) when 1 week of $200 subscription gets you $800~$1200 sure is something.
> I can just pull out the Amex and everything works
Must be nice to have money to burn. But if I had to guess, this doesn't seem like representative consumer behavior that Anthropic or OpenAI should rely on holding up at scale.
My theory is that for a large amount of people, “work are paying anyway so what do I care?”.
I lost my job recently, so gave myself an AI budget of $100 to help with search and applications.
If I put that in OpenAI or Anthropic, I’d hit limits quickly and lose whatever I didn’t use.
Or… I could put the same money in OpenRouter, use open models at 1/25th the price, and only need to pay more when I’ve spent what I put in.
Back in January when Claude Cowork was new and Claude Code was one of the only performant harnesses, it would have been a tougher decision.
But Hermes, dsh, agy, codex, opencode, pi… they make it so easy to achieve so much with such a low budget.
I know it’s a cliche nowadays to say “just use cheaper models” but the value they offer is SO much greater. And i can switch to GPT6, or Opus 5.5 in two clicks for tasks anyway.
My point is this: OpenAI and Anthropic are pulling stunts like this because they don’t care about you and your subscription. So stop caring about them.
> My point is this: OpenAI and Anthropic are pulling stunts like this because they don’t care about you and your subscription. So stop caring about them.
Exactly where I landed, I had a personal subscription last year that shifted between OpenAI and Anthropic depending on who had the better model for that month.
This year? I don't need that, I can do the same as you: budget and pre-pay for some tokens in OpenRouter, and use very cheap models for absolutely anything I need on my personal projects. I can use open source harnesses that give me similar results, my projects do not need the absolute most-expensive frontier model at all, that's just a waste.
And if absolutely needed to use some frontier capability I can pay the tokens for that instead of committing to US$ 200-500 for a subscription that they can just pull the rug from me at any point.
I still have my job where they give me access to all the shiny expensive models with their enterprise agreements about data retention, the legal stuff that a company cares about and my personal projects don't, if I keep my usage under the newly implement monthly budget no one will bother me about it and so I just use what I'm told to.
Ideally you use so many different compute providers that your footprint isn’t as complete for any of them. If you commit to a single party they have total information which is more dangerous. If you’re into that sort of thing. Obviously just don’t put anything in a http request you don’t want anyone else to know…
One of their three identical accounts had ~20% lower limits and the explanation was an "extremely tiny" A/B test. Makes the "5x" kind of a joke. I'd honestly take a slightly worse plan if it just published the actual token limits.
They would never do this because it would allow customers to see how often and to what degree they're shifting account limits. Token transparency costs extra, full API price to be exact.
The title is misleading: by token count, it’s only 3× more valuable.
And there’s also a major error in the calculation: it doesn’t account for the “generous” resets that are mentioned at the beginning of the article!
Taking the 1 to 2 weekly resets into account and comparing tokens, the ChatGPT Pro 200 plan works out to be between 1.2× and 1.7× more cost-effective per token than Claude’s for Astra/Fable, while for Sol/Opus it’s more like 0.7× to 1.04× as valuable.
However, this also doesn’t take into account the fact that, as another commenter pointed out, GPT-6.1 Sol currently uses fewer tokens to perform the same task (according to that comment, 5× fewer tokens, but I haven’t verified those numbers).
My figures are very rough, but the conclusion in the title is clearly false and clickbait. I get the impression that OpenAI offers better value for now, even if that means we’re relying on those “generous” resets continuing to happen.
I’d agree if Claude’s limits weren’t changing and getting lower month after month. For now, neither one is profitable for their lab, and I don’t find their plans very predictable.
> I want a predictable LLM inference subscription, not a gacha game.
The only way we get to predictability is by paying what subscriptions are actually worth, and this entire game hinges on the fast that AI companies are not convinced that most people are willing to pay the 3-5x (or however much it is) multiple on what these accounts actually cost to run.
It's an advertisement. Just like 90% of the AI posts on here and all of social media. I bet it's less than 3x in reality too. Especially once anthropic needs to become profitable for their IPO...
It sure looked to me like they were claiming it was an advertisement for Anthropic. (And in any case I don't think an article becomes an advertisement merely because you sell collections of similar articles.)
I have both, 6.1 sol is no where near 5x more efficient then opus in day2day. Without numbers i would not even entertain that though and with numbers i would look closly if its the same code. I put it to work on the same repo so i have somewhat of a comparison. Not to mention that opus 5.5 is way better in most of the tasks. Currently openai gives you less for worse idk how did they think that could work.
> Token efficiency is also an extremely relevant factor, but the industry unfortunately lacks reliable data here. Many people like to cite this chart from Artificial Analysis, but we do not believe the benchmark tasks in the AA Intelligence Index are at all representative of real work people do with LLMs.
Right, so they just ignore the fact that different models use very different number of tokens to achieve the same thing. Ignoring that means they can't say anything about "value". This is like comparing numbers with different units. They're Atokens and Otokens.
if anyone wants a lifehack just get $100 sub from anthropic, get $80 sub from z.ai, download omp.sh, set task/advisor to glm 5.3 flash with max thinking, default opus 5.5 with medium thinking, slow set to xhigh thinking and design set to xhigh/max - max will produce significantly more complete designs.
day2day performance difference is negligible, there's less refusals and you always have glm 5.3 for whenever opus 5.5 arbitrarily decides to terminate your conversation.
This is enough to run 2 concurrent sessions 16 hours a day.
Are you using a burner number for each one? I'm surprised they don't ban by phone number and/or payment method. Or that you haven't been hit by identity verification.
Somewhat tuning out these "X is better than Y" news and articles, as things are still moving quite fast. In a few weeks, OpenAI offers more value for money and the cycle repeats again.
As a Claude subscriber, I am quite enjoying the value I get from Anthropic. But I am not forgetting that this is their loss leader, and API access is still their main driver. So strike while things are good, but expect some pullback in the future.
Hmm I'm not sure this is really the case I have the pro level of each, and GPT 6.1 Sol feels virtually unlimited to me atm, doesnt feel the same with opus and its probably only a few percentage points better and not in all situations.
I have subscriptions to both, one for personal and one for work use, and I haven't really come close to hitting the limits since the release of Opus 5.5 and GPT-Sol 6.1, and I'm no longer hitting any of the capability limits I used to experience when using the "mid" tier models, so I don't miss using Astra. It feels like limits at pretty much a solved problem right now, for my level of use.
I'm absolutely hitting the limits with 6.1. I'd say maybe hitting them a bit slower than before, but only because the model is dog slow, not because it's more efficient.
What’s news to me is that Anthropic reports that it has 88% gross margin. I find that extremely hard to believe with the capex required to build and run the models.
Accounting 101: CapEx doesn’t directly affect either gross or net margin (when the expenditure occurs). It is capitalized on the balance sheet and affects profitability over time (through depreciation or amortization).
It’s not really over time though is it - it recorded a net loss of $42billion this year. So one can sandbox these figures all one wants and show such profitability but even a casual look at the numbers shows that 88% figure to be a nonsense. But ok fine accounting yadda yadda they’re profitable.
I wonder how subscription breakage (people who pay, but don't use a fraction of their subscribe-for-paid-token-count) plays into the frontier labs gross margins. A number we will likely understand if/when they go public and something users can track locally if they choose.
How do we know 100% subscriptions are subsidized, it’s a possibility that they are actually ok (ie some people over use, majority under use and on net profitable) and the API pricing is actually just super expensive. Subscription is a buffet and API is a private chef :)
I am sure we will all be using local models or GPU hosted models of our own soon. The prices of frontier labs is bordering on premium for the value they provide when compared to the open models I use.
But back in February the OpenAI offer was better. Have both of them. Nowadays Anthropic really is the best again. But only after DeepSeek raised its prizes.
This says “Anthropic gives you 5× more API-priced model usage”, which is not equivalent to “Anthropic is 5x as much value”. This does not take into account model/harness efficiency. I tried looking up some benchmarks and Claude seemingly uses roughly 5x as many tokens while taking its sweet time to complete a task. No such thing as a coincidence.
My chatgpt pro account gave me 62k of their "credits" which they try to sell you at extortionate rates when your subscription runs out. I spent two days hammering it and used about 4k of those.
This is clear manipulation to prevent me from complaining about lower limits until it's out of the news cycle, but I don't care. As soon as those are gone, if my $200 account doesn't give me the value it used to, I'm out.
The random resets are also clear manipulation in this regard as well. Rather than just give me a fixed weekly limit which I could then get a sense of and know if it's getting reduced, they throw out a ~0-3 resets a week at random, unexpected intervals so that I'll never know.
I don't think it's them exclusively.I'm the admin on the claude account at work and I have their cheap account and never hit my limit, while others are on the expensive plan and run through hundreds of dollars of credits in addition each month. I'm not taking it easy on this thing either. I had it convert several thousand rules in our firewalls to fortimanager global rules, I gave it a personal project with a long promp to build a truenas replacement based on universal blue's ucore, webzfs and config files for nfs, cifs/samba, and s3 (rustfs) all managed through my forgejo setup with forgejo runners building it and forgejo providing the repo that the machine looks at to upgrade the base and the config (it's two seperate pathways so I can push config changes like new nfs shares without having to do a system image update). All that barely moved the needle on my account. I suspect they are all a bunch of shysters opportunistically abusing us.
> I suspect they are all a bunch of shysters opportunistically abusing us.
At this point we conclusively know that.
What frustrates me is the constant unshakeable belief here that subscriptions must be running at a loss, based on zero evidence.
Anthropic have said they have 85% margin, which presumably means their normal API pricing. 85% margin is insane. No business deserves that. We can easily assume the subscription is running at a more sane margin like 5-15%.
What margin does a business “deserve”? If people are willing to pay it, what do you expect to happen?
They deserve nothing. They take what they can take and government antitrust needs to be there to break up market failures and that is about it.
I would pivot to a Chinese provider via open router so fast... Most months on glm or DeepSeek it would be a challenge to spend $100. But now I mostly use local models... Sounds obnoxious bouncing between 2 companies who rug pull their customers every 4 months and degrade service as a form of advertising...
> But now I mostly use local models
Do you have beefy hardware? i dream of using a local LLM, but neither my 24GB RAM M4 Air nor my PC with a RTX3080 and 16GB of RAM seem usable (yet).
i could label them usable if "prompt it and then wait for 6 minutes for it to change a single line" is deemed usable, which i don't.
spent the last weekend in a rabbit-hole of which models to use in what kind of setup, but i think with my hardware i'm just out of luck for now.
for private matters I can use Cloud AIs, but not allowed to use it for work-related matters, which is where i could use it the most. they do provide us an isolated AI environment where we can use Claude etc. for work stuff, but heavily rated (about 20 prompts per week per Claude model).
You're super close on your PC. It seems like the sweet spot right now is 24gb vram + lots of RAM for qwen 3.8 flash next or a dgx spark like machine with buckets of unified RAM.
You could run some decent models on your PC l, but nothing that would totally replace serious work. Qwen 3.7 35b is pretty okay. Maybe you could try the new bonsai quant for 3.8 27b but I doubt it would be awesome.
i did in fact test different Qwen's! :) qwen3.5:9b and qwen2.5-coder:14b, i think. but i struggled with speed and them being stuck in "Hmm, wait, no, that doesn't work, let me analyze again"-loops. on the M4 Air I was able to start some bigger models. definitely the first time i burned through 25% of battery in half an hour while getting first-degree burns on my lap.
maybe in 1-2 years my hardware will be enough for fast local-AI. maybe by then i can afford new hardware. (think with current trends option #2 gets less realistic each quarter).
not sure if i lost a bit of curiosity and spark. there's so many models, tools, harnesses, tweaks for harnesses, edited models and so many things. ultimately i don't care enough, it seems vapid to stay bleeding-edge informed about LLMs, if one's career does not directly depend on it. i mean, what the fuck even is a bonsai quant, this stuff makes me feel like an uninformed excel boomer even tho i absolutely don't am one :D
not a software engineer per se so i don't need 16 agents running 24/7 with openclaw. i want a local buddy who helps me write ansible/python faster, better, helps me analyze bugs, refactors small things. basically what Claude Web does, with the benefit of it reading the files itself and running locally.
maybe i just ignore the whole AI/LLM part and just buy some new hardware so i can use higher graphics settings while gaming, WAIT NO, that market was nuked by AI as well, guess i'll continue using my i5 from 2019.
> i mean, what the fuck even is a bonsai quant, this stuff makes me feel like an uninformed excel boomer
Local models and harnesses are the new ~2010s js front end frameworks.
Either accept there's always going to be a new thing of the week and don't over-invest in expecting reliability, or work in another space.
I know it's anathema to "normal" reliable software dev, but that's what new industries look like...
[flagged]
Do you need the $200 sub to use the credits?
As soon as I received them and realized they expire this year, I switched to a $20 membership (billing cycle starts tomorrow), assuming you don't.
After consuming them, I will decide to whom subscribe next.
I did the same and my billing also starts tomorrow so... I guess we'll find out!
afaik you can even cancel your sub and run entirely on those credits :)
They also carefully sequence the resets so they bunch up all at once, including with the regular weekly reset.
Last week I've have about 4 resets in 48 hours:
- reset done by OpenAI around Friday
- banked reset expiring on Saturday
- banked reset expiring on Sunday
- regular weekly reset expiring on Sunday
So basically I was unable to use them, given weekend and all at once, unless you have the software factory ready to spin up...
I'm not complaining, they are free after all, but it's clear they are not randomly distributed.
> subscription gross margins are way lower than API and meaningfully reduce revenue per MW for both OpenAI and Anthropic
I don't know about everyone else, but there is no way in hell I would pay even $200 for API-priced tokens for personal use. So at least for my sample size of one, Anthropic's revenue would not be higher if they dropped their subscription plan (they would get $0 from me instead of $200 per month).
I’m not everyone else but I am SOMEBODY else and you’re absolutely right. I was paying for Claude since they first offered subscriptions, right up until this summer.
I respect Anthropic (and OpenAI to a lesser extent) but I’m not going to play these games.
$200/mo for hiring some very capable interns, that's not very expensive.
And a local AI solution is less capable and also quite expensive, but for some worth trying.
Local ai can be about as capable as anything you get a subscription for. The benefit to local beyond a lot of other things is, no model degradation, and stationary costs. Actually, you get effectively free model upgrades.
For 2-4k you can have opus 4.8 at home running faster than anthropic. In 8-20 months you have broke more than even.
I think for almost anyone it's worth trying. Especially if you already have hardware.
It's not "about as" capable. It might be capable enough, depending on use case however.
My use case is programming. This is a viable replacement for me.
Right, but saying local models are just as capable frames it like they're equal. Which is not the case
Yes, but we are getting close to the point where the difference exists, but it is less and less significant. Tbh, Qweb27b is perfectly find for a lot of tasks; anecdotally, the difference lies less and less in quality of responses, but in speed. Local models have a hard time there for a reason ( the cheapest stuff you can get now is stuff like previously mentioned 128gb ryzen ai ).
So trade-offs are there, but do those matter to all people the same way ( are they not equal in the same way to everyone )?
I would say from experience, that yes the local models today are as good as the old frontier models. That was my original claim. I am sticking too it. It won't be long before local models match say fable 5.x as well. And people will still say "it's not as good as x". No it's not but it's as good as the thing everyone freaked out about 6 months ago. If that wasn't good enough for all y'all then then why did you all use it? Exactly.
This is the elephant in the room that OpenAI and Anthropic specifically aren't talking about -- when "good enough" is surpassed.
Because there are a large number of use cases (and large sub-portions of others) where genius-level AI isn't required.
Which means once that threshold is surpassed in people's relevant fields, available margin on that is going to collapse to commodity levels.
It's difficult to see how either pure-play AI company maintains its valuation once that happens. Their TAM is based on capturing a big chunk of all work, not just that which requires the highest intelligence.
$4k doesn’t even buy you a DGX spark these days. You can get a Ryzen AI max 128GB machine for a bit less than $4k but the models you can run on there aren’t even close to Opus 4.8.
Amazon recently moved $8B in GPUs to a SPV to hide away their expenses. Oracle warned their data center isn’t going to be on schedule. Worst of all, none of the big players are reporting enough deprecations to match up with GPU sales.
All this indicates to me that loads of GPUs were only purchased on paper or are sitting in warehouses unused. This almost certainly means a drop in orders followed by a drop in RAM prices (though that would indicate a demand drop to investors, so maybe it’s better for stock prices to keep paying to bills bigger warehouses and stuff them full of unused GPUs).
The second RAM prices normalize, local LLM becomes much cheaper. A machine that doesn’t make much sense at $10-15k suddenly becomes a lot more feasible at $3-5k. If those stored GPUs flood the market, we might see even bigger price drops.
that's a wild theory. You think there are really warehouses of GPUs that they are just sitting on until data centers are built, that will just all of a sudden flood the market?
First, the form factor of GPUs in data centers aren't the same as desktop GPUs, so you couldn't use them even if you wanted to in a normal rig.
Second, from a business perspective that doesn't make a lot of sense. It's much more likely that GPUs are going to the highest bidder/large contracts who are scooping them up to populate/upgrade data centers that are in operation because they are going to get more money per gpu on newer hardware. The "old" hardware might go to a warehouse to be auctioned to the highest bidder or go to a data center coming online, but I highly doubt they are sitting on market wrecking amounts of GPUs just waiting to flood the market. When they go bankrupt and they have to sell a datacenter or two, those datacenters being parted out as part of a bankruptcy deal I could see. Warehouses full of unused GPUs doesnt make any sense to me though.
> First, the form factor of GPUs in data centers aren't the same as desktop GPUs, so you couldn't use them even if you wanted to in a normal rig.
With eGPU PCIe 3.0 or 4.0 x4 links (or even Thunderbolt), that essentially doesn't matter for inference.
You pay the bandwidth hit on model load / unload (a few seconds), but it's irrelevant for post-loaded inference.
Anyone hosting a high wattage GPU is going to be fine hosting in an external enclosure, most of which have generous extra-spec room.
And if they don't, if a flood of cheaper DC GPUs hit the used market, you can bet Chinese manufacturers will have enclosures that fit them available the day after.
"Microsoft Is Warehousing An Estimated $50 Billion to $100 Billion In GPUs" https://www.wheresyoured.at/wherere-all-the-ai-chips/
It makes sense if you consider that maybe all of these companies are colluding to forward a shared agenda. It is happening whether it makes sense or not. So there's that too.
>>The second RAM prices normalize, local LLM becomes much cheaper. A machine that doesn’t make much sense at $10-15k suddenly becomes a lot more feasible at $3-5k. If those stored GPUs flood the market, we might see even bigger price drops.
Yeah but it's unlikely to happen in coming 2-3 years (at least) and then the SOTA models are going to be a completely different animals.
On a single 3090 while offloading experts to RAM you can run qwen 3.8 next. Which yea is about 2.5k new and yep about opus 4.8.
I would love for this to be true, but this is nonsense. The local models that you can run for those prices are not as capable as current frontier. At $200 a month, it would take well over a year to break even, and that’s if you don't include energy prices and time spent tinkering to get and keep it working.
I exclusively pay for API tokens. This is the primary way I consume these models.
The flat monthly subscriptions seem too tempting for the providers to screw with. I prefer to paygo and to be responsible with my consumption. I also want the ability to scale substantially beyond what a typical consumer plan may offer on occasion.
My monthly usage ranges from $10-$1000. I don't have to worry about quotas or anything. If I need to use several thousand dollars worth of tokens, I can just pull out the Amex and everything works. It's constant performance all day every day. I have long since maxed out my org level with OAI, so it would be very difficult to exceed any realistic limits.
Spending 5/10 months equiv of subscription cost that have significantly lower revenue for X company (or even costs money) when 1 week of $200 subscription gets you $800~$1200 sure is something.
What's stopping you from using the subscription plan and then topping up with API if you need?
> I can just pull out the Amex and everything works
Must be nice to have money to burn. But if I had to guess, this doesn't seem like representative consumer behavior that Anthropic or OpenAI should rely on holding up at scale.
My theory is that for a large amount of people, “work are paying anyway so what do I care?”.
I lost my job recently, so gave myself an AI budget of $100 to help with search and applications.
If I put that in OpenAI or Anthropic, I’d hit limits quickly and lose whatever I didn’t use.
Or… I could put the same money in OpenRouter, use open models at 1/25th the price, and only need to pay more when I’ve spent what I put in.
Back in January when Claude Cowork was new and Claude Code was one of the only performant harnesses, it would have been a tougher decision.
But Hermes, dsh, agy, codex, opencode, pi… they make it so easy to achieve so much with such a low budget.
I know it’s a cliche nowadays to say “just use cheaper models” but the value they offer is SO much greater. And i can switch to GPT6, or Opus 5.5 in two clicks for tasks anyway.
My point is this: OpenAI and Anthropic are pulling stunts like this because they don’t care about you and your subscription. So stop caring about them.
> My point is this: OpenAI and Anthropic are pulling stunts like this because they don’t care about you and your subscription. So stop caring about them.
Exactly where I landed, I had a personal subscription last year that shifted between OpenAI and Anthropic depending on who had the better model for that month.
This year? I don't need that, I can do the same as you: budget and pre-pay for some tokens in OpenRouter, and use very cheap models for absolutely anything I need on my personal projects. I can use open source harnesses that give me similar results, my projects do not need the absolute most-expensive frontier model at all, that's just a waste.
And if absolutely needed to use some frontier capability I can pay the tokens for that instead of committing to US$ 200-500 for a subscription that they can just pull the rug from me at any point.
I still have my job where they give me access to all the shiny expensive models with their enterprise agreements about data retention, the legal stuff that a company cares about and my personal projects don't, if I keep my usage under the newly implement monthly budget no one will bother me about it and so I just use what I'm told to.
> Or… I could put the same money in OpenRouter, use open models at 1/25th the price
But you give up privacy, because now your data is in the hands of 2 service providers, not 1.
Ideally you use so many different compute providers that your footprint isn’t as complete for any of them. If you commit to a single party they have total information which is more dangerous. If you’re into that sort of thing. Obviously just don’t put anything in a http request you don’t want anyone else to know…
> Ideally you use so many different compute providers that your footprint isn’t as complete for any of them.
Hmm,yeah but then come the data brokers who conveniently tie everything back together ...
i am partial to qwencloud
One of their three identical accounts had ~20% lower limits and the explanation was an "extremely tiny" A/B test. Makes the "5x" kind of a joke. I'd honestly take a slightly worse plan if it just published the actual token limits.
Grab a plan from one of the Chinese companies, they give you explicit ranges token that you can expect to get.
They would never do this because it would allow customers to see how often and to what degree they're shifting account limits. Token transparency costs extra, full API price to be exact.
The title is misleading: by token count, it’s only 3× more valuable.
And there’s also a major error in the calculation: it doesn’t account for the “generous” resets that are mentioned at the beginning of the article!
Taking the 1 to 2 weekly resets into account and comparing tokens, the ChatGPT Pro 200 plan works out to be between 1.2× and 1.7× more cost-effective per token than Claude’s for Astra/Fable, while for Sol/Opus it’s more like 0.7× to 1.04× as valuable.
However, this also doesn’t take into account the fact that, as another commenter pointed out, GPT-6.1 Sol currently uses fewer tokens to perform the same task (according to that comment, 5× fewer tokens, but I haven’t verified those numbers).
My figures are very rough, but the conclusion in the title is clearly false and clickbait. I get the impression that OpenAI offers better value for now, even if that means we’re relying on those “generous” resets continuing to happen.
No serious analysis can consider the completely discretionary resets. I want a predictable LLM inference subscription, not a gacha game.
I’d agree if Claude’s limits weren’t changing and getting lower month after month. For now, neither one is profitable for their lab, and I don’t find their plans very predictable.
We should take advantage of it while it lasts.
> I want a predictable LLM inference subscription, not a gacha game.
The only way we get to predictability is by paying what subscriptions are actually worth, and this entire game hinges on the fast that AI companies are not convinced that most people are willing to pay the 3-5x (or however much it is) multiple on what these accounts actually cost to run.
It's an advertisement. Just like 90% of the AI posts on here and all of social media. I bet it's less than 3x in reality too. Especially once anthropic needs to become profitable for their IPO...
> It's an advertisement.
Evidence?
They may be referring to the fact that semianalysis sells this information for $500/yr newsletter.
It benefits them to have sensational headlines that people might pay to have earlier/better/more of.
It sure looked to me like they were claiming it was an advertisement for Anthropic. (And in any case I don't think an article becomes an advertisement merely because you sell collections of similar articles.)
I have both, 6.1 sol is no where near 5x more efficient then opus in day2day. Without numbers i would not even entertain that though and with numbers i would look closly if its the same code. I put it to work on the same repo so i have somewhat of a comparison. Not to mention that opus 5.5 is way better in most of the tasks. Currently openai gives you less for worse idk how did they think that could work.
Also, bear in mind the guy who wrote the article lives in the same flat as Sholto Douglas, an Anthropic researcher.
More confusingly, they both use different tokenizers so token counts aren’t 1:1 at all.
[flagged]
This is a joke.
> Token efficiency is also an extremely relevant factor, but the industry unfortunately lacks reliable data here. Many people like to cite this chart from Artificial Analysis, but we do not believe the benchmark tasks in the AA Intelligence Index are at all representative of real work people do with LLMs.
Right, so they just ignore the fact that different models use very different number of tokens to achieve the same thing. Ignoring that means they can't say anything about "value". This is like comparing numbers with different units. They're Atokens and Otokens.
if anyone wants a lifehack just get $100 sub from anthropic, get $80 sub from z.ai, download omp.sh, set task/advisor to glm 5.3 flash with max thinking, default opus 5.5 with medium thinking, slow set to xhigh thinking and design set to xhigh/max - max will produce significantly more complete designs.
day2day performance difference is negligible, there's less refusals and you always have glm 5.3 for whenever opus 5.5 arbitrarily decides to terminate your conversation.
This is enough to run 2 concurrent sessions 16 hours a day.
These prices are getting ridiculous, we’re talking about 180$ / month here.
If you're actually running 2 sessions 16 hours a day, I would hope you are getting $180 of value. (also, maybe work less)
1 hour of my time costs more than that.
Is anthropic allowing 3rd party harnesses like that?
No, this risks getting you banned.
just make a new account I am on my 16th or something, but they don't really ban anymore as I have had this one for like 8 months.
Are you using a burner number for each one? I'm surprised they don't ban by phone number and/or payment method. Or that you haven't been hit by identity verification.
smspva + my bank allows me to have infinite payment cards
Interesting. Any haven't been hit by identity verification?
What happened with your subscription when they banned you? Do you lose what's left of the month without a refund?
they just refund you and they usually ban you within a week, but as I said they kind of just stopped or largely have given up.
They ban by payment method, but omp.sh mimics the Claude Code useragent + payload, so the risks are a lot lower.
Somewhat tuning out these "X is better than Y" news and articles, as things are still moving quite fast. In a few weeks, OpenAI offers more value for money and the cycle repeats again.
As a Claude subscriber, I am quite enjoying the value I get from Anthropic. But I am not forgetting that this is their loss leader, and API access is still their main driver. So strike while things are good, but expect some pullback in the future.
Hmm I'm not sure this is really the case I have the pro level of each, and GPT 6.1 Sol feels virtually unlimited to me atm, doesnt feel the same with opus and its probably only a few percentage points better and not in all situations.
I have subscriptions to both, one for personal and one for work use, and I haven't really come close to hitting the limits since the release of Opus 5.5 and GPT-Sol 6.1, and I'm no longer hitting any of the capability limits I used to experience when using the "mid" tier models, so I don't miss using Astra. It feels like limits at pretty much a solved problem right now, for my level of use.
I'm absolutely hitting the limits with 6.1. I'd say maybe hitting them a bit slower than before, but only because the model is dog slow, not because it's more efficient.
What’s news to me is that Anthropic reports that it has 88% gross margin. I find that extremely hard to believe with the capex required to build and run the models.
Accounting 101: CapEx doesn’t directly affect either gross or net margin (when the expenditure occurs). It is capitalized on the balance sheet and affects profitability over time (through depreciation or amortization).
[1] https://en.wikipedia.org/wiki/Capital_expenditure
It’s not really over time though is it - it recorded a net loss of $42billion this year. So one can sandbox these figures all one wants and show such profitability but even a casual look at the numbers shows that 88% figure to be a nonsense. But ok fine accounting yadda yadda they’re profitable.
I wonder how subscription breakage (people who pay, but don't use a fraction of their subscribe-for-paid-token-count) plays into the frontier labs gross margins. A number we will likely understand if/when they go public and something users can track locally if they choose.
How do we know 100% subscriptions are subsidized, it’s a possibility that they are actually ok (ie some people over use, majority under use and on net profitable) and the API pricing is actually just super expensive. Subscription is a buffet and API is a private chef :)
This is a SemiSeriusAnalysis. The value I get isn’t about the discount rate of tokens, it’s About the amount of work I can get done.
There lack of accounting for token efficiency and tokens/task. Makes this whole thing are less useful.
[flagged]
I am sure we will all be using local models or GPU hosted models of our own soon. The prices of frontier labs is bordering on premium for the value they provide when compared to the open models I use.
> but the gap is still massive even if you switch to comparing the number of tokens.
Token doesn't have the same meaning at OpenAI vs Anthropics, they use different tokenizer. How can this be used for comparison?
But back in February the OpenAI offer was better. Have both of them. Nowadays Anthropic really is the best again. But only after DeepSeek raised its prizes.
This says “Anthropic gives you 5× more API-priced model usage”, which is not equivalent to “Anthropic is 5x as much value”. This does not take into account model/harness efficiency. I tried looking up some benchmarks and Claude seemingly uses roughly 5x as many tokens while taking its sweet time to complete a task. No such thing as a coincidence.
Also how tokenization works. There was a post the other day saying 150k OpenAI tokens is equivalent to 250k Anthropic tokens.
Can't find the link now it was a the pi harness author saying you could load the entire harness into context with those budgets IIRC.
Let's hope they don't nerf Opus 5.5
im legit praying this doesn't happen
There's a simple explanation here.
Anthropic has way more inflated api token priced than OpenAI what is plainly visible on cost charts.
Submarine ?
but ai credit runs out really really fast