That's malicious and I think this is scamming from the literal money (you didn't do anything wrong, you executed one command and they scammed you out of the fair usage you paid for).
Please raise the ticket or at least GitHub issue for visibility.
Sooner or later some sort of complaint to the relevant trade authority should happen - this is a scam operation at this point.
There are many possible explanations for this outcome to have occurred other than malice. If you're an engineer by trade, consider how many bugs you've been responsible for over the course of your career that you didn't intend. Probably a lot.
There's been a sustained pattern of incidents. If Anthropic were truly serious about not wanting to take people's money, then they would have put in place whatever review processes were necessary to stop this from happening. So regardless of whether or not they specifically intend to cause harm, they're willingly letting it happen, which is just about as bad.
Yes, it's reasonable to turn down the heat. But it's also reasonable for people to be upset when their money is taken from them, and when the company that does so is effectively beyond persecution for doing so.
Even with the best of faiths, this is at the very least a shoddily vibe coded “detect and low-key block attempts to use Claude for Openclaw” - it decided to look for specific strings wrapped in json without realizing this doesn’t always imply it’s an actual payload for Openclaw itself. And the human driving it was too dumb to review/catch this bad inplementation.
So maybe not malice, but certainly a level of ineptitude I don’t expect from a crucial vendor from a tool that’s become essential for many developers.
(I don’t care, I do just fine when Claude is down or refuses to help me (it has happened) though)
Also they ain't wrong. In what other context does OpenClaw get mentioned?
"You may not use our service if you mention OpenClaw" is a harsh line but hardly illegal or forbidden any more than any other service restriction (i.e. no use allowed for high-stakes financial modeling). Don't like it, cancel your plan.
But that's the thing -- there is no line! Where is this specified? How can we know what service restrictions there are? For all I know, my plan could be exhausted at any point during the workday just because I happened to touch on some keyword Anthropic has decided to ban.
> Don't like it, cancel your plan.
Ah, but I thought these models were supposed to have been trained for the sake of humanity? That the arbitrary enclosure of the collective intelligence was for our own good? These concepts are not compatible.
All of this is just criminal and fraudulent behavior, done July a whole bunch of people who haven't learned their lesson, and keep sending Anthropic more money for abuse at scale.
It's in the TOS, so no, not fraud. You might not like it that Anthropic doesn't want you running OpenClaw (effectively owned by a competitor) on CC, but that doesn't make it fraudulent or criminal.
There is literally nothing close to illegal about this behavior. You read the terms of service right, which provides a long list of explicit and implicit disclaimers?
That's rather shitty. It's one thing to disallow bypassing preferential pricing models, it's a completely different thing to castrate your model against some uses.
You can see how it goes in the future. Wanna vibe code a throwaway script? $0.20. Ah, it's for a legal document search? $10k then. Oh and we'll charge 20% of your app sales too - I can see how they are going in real time, mind you!
The question is how much friction there will be for people to switch over to Gemini, GPT or maybe even DeepSeek or Mistral or whatever. Even if price hikes are inevitable across the board, the moat any single org has is somewhat limited, so prices definitely will be a factor they'll compete on with one another at least a bit.
This is assuming there will be no competition. But why wouldn't there be? Especially since you can use open source models, which are not too far from frontier models (from now).
Deepseek has demonstrated that there is no reason for it to actually lose money. The awful business practices and monopoly tactics of the frontier model labs in the US are the problem.
I think it goes beyond this. I was just using claude to edit a blog post which mentioned OpenClaw and I got this response: "The "OpenClaw" reference — I assume that's a typo or playful reference; if you mean a real product, I couldn't find it under that spelling and you'll want to fix or footnote it.". I gave it a direct link to openclaw.ai and the chat instantly ended and hit my 5hr usage limit. Could have been a coincidence, but I had only lightly been using sonnet in the morning so it seems unlikely. Very odd.
Dragons steal gold and jewels... and they guard their plunder as long as they live... and never enjoy a brass ring of it. Indeed they hardly know a good bit of work from a bad, though they usually have a good notion of the market value
The weird thing is that it found sources for all of my other claims and references no problem, but acted like it didn't know what openclaw was when openclaw.ai is the first thing that pops up on google.
It also sounds extremely counterproductive to try and sabotage your competition by.. driving your customers away? I have no love for these companies but it's a silly conclusion to jump to.
Not if a chatbot did it, maybe. No legal precedence here. Also they are a defense and offense contractor they could kill people and nothing would happen
Claude.ai is now at a 98.85% uptime. There's been so many frustrations with Claude / Anthropic lately (very heavy usage limits, wrong A / B testing, etc.).
I have been really happy with my Codex subscription lately, but feels like these things change every other day. The OpenCode Go subscription for trying out GLM, Kimi, Qwen, Deepseek and friends also looks useful.
But nonetheless, Opus 4.6 is a very capable model, but justifying a Claude subscription gets more and more difficult, think I might just sometimes use it through OpenRouter or as part of something like Cursor (although I'm not sure about the value of that subscription as well).
There were periods where I was entirely unable to use Claude Code for hour+ due to auth gateway always returning 500 or timing out, there was an "elevated errors" incident shown on status.claude.com, but zero minute of downtime recorded (not even "partial outage"). So the real uptime should be even worse.
April has been a crazy month for open weights models. I've been using Claude Code for work and Kimi 2.6 for personal projects and Kimi has been very good. Glm-5.1 is also great. Qwen, Mimo and Deepseek I need to test some more, but they all have been producing good results. I have the impression that they are all are at the same level, or close to, Sonnet 4.6.
The little qwen36 is at sonnet level . Kimi2.6 is about opus. The one can run on a single GPU on your gaming pc. The other you can run way cheaper from a provider. Or if you are really wealthy and have lots of gpus can run it yourself.
Would "lots of gpus" even help for huge models? Maybe this is exposing my lack of knowledge but don't you need to keep the whole model and context in a single GPU's VRAM? My understanding is that multiple GPUs help with scaling (can handle N X inference requests simultaneously) but it doesn't help with using large models. If that were the case, I could jam another GPU in my box and double the size of model I can serve.
> Would "lots of gpus" even help for huge models? Maybe this is exposing my lack of knowledge but don't you need to keep the whole model and context in a single GPU's VRAM?
How do you think the large providers do inference? No single GPU has 1TB plus of memory on board. It’s a cluster of a bunch of gpus.
Please don't oversell them. Eg Kimi k2.6 has a maximum context size of 270k, that's a quarter of opus.
The model is fine, Ive switched to it entirely for a personal project, but it's not opus.
And no, you're not running then locally unless you're a millionaire. You still need hundreds of GB (500+++) of VRAM on your graphics card - that's not at a level of consumer electronics.
Sure you can run the quantized models, but then you're at Haiku performance.
This is very concerning. Their heavy handed tactics haven't impacted me personally yet but I am increasingly nervous and casting about for viable egress paths if I need to flee Claude Code. I really hope they pump the breaks and thoroughly reorient themselves. They are under a lot of competing pressures and probably can't make a decision that won't upset a lot of people (in order to balance growth and capacity etc), but are coming to the worst possible conclusions.
For instance, maybe you can't afford to take on more customers right now, Anthropic. Maybe if you are severely undermining the customer relationships you already have, you should just admit you can't sell any more 20x plans right now and only accept new customers at lower tiers until you have the necessary capacity.
This is also a DoS you could drive a truck through, and it's disturbing such an obvious vulnerability was shipped at all.
> or instance, maybe you can't afford to take on more customers right now, Anthropic. Maybe if you are severely undermining the customer relationships you already have, you should just admit you can't sell any more 20x plans right now and only accept new customers at lower tiers until you have the necessary capacity.
Or just increase prices for new claude code users? Surely transparent upfront across the board price increases are easier to swallow than hidden context-based pricing changes like this?
> casting about for viable egress paths if I need to flee Claude Code
Check out OpenCode (the OSS product [1]) and OpenCode Go/Zen (the LLMaaS [2]). Use a more expensive model with larger context (like GLM-5.1) for orchestration and cheaper models for coding and iteration on acceptance criteria (writing and passing tests). I also throw a more expensive vision-capable model into the mix like Gemini 3 Flash to iterate on UI tasks using Playwright. With the base usage in Go and pay as you go on cheaper models like MiniMax you can get a lot done for not a lot of coin.
They have better PR than OpenAI but they are not a more ethical company. They do a bunch of shady stuff and are just as much involved in military applications. Cal Newport’s recent podcast had a good discussion about this: https://youtu.be/BRr3pAPsQAk?si=jaRJYJ_XQE7VpxPN
I understand not everyone has the interest or time to sit through an hour long podcast. But last I checked this is HN, and I think that podcast is right up the alley for many of us here. Cal Newport is not exactly a 'random podcaster'.
Next time I can summarize some of the talking points in my comment though, but I didn't want to poorly regurgitate the arguments when they were readily available in the video lol.
Although I see another poster has commented the key takeaways :)
Cal Newport and tech commentator Ed Zitron discussed this disparity between Anthropic's public image and their actual practices. Despite cultivating a reputation as the "ethical" AI company, Zitron argues that Anthropic's actions show they are just as ruthless and ethically questionable as their competitors.
Anthropic has been deeply integrated with the US military, having been installed with classified access since June 2024. The podcast highlights that Claude has been actively utilized during the "Venezuela incursion" and the ongoing "war in Iran".
Despite this active involvement, CEO Dario Amodei released a statement attempting to publicly distance the company from the Department of Defense by declaring they would not allow their technology to be used for "mass domestic surveillance" or "fully autonomous weapons". Zitron categorizes this as a highly calculated PR maneuver, pointing out that LLMs are fundamentally incapable of controlling autonomous weapons anyway. The stunt successfully manufactured a wave of positive press—with celebrities and commentators praising Anthropic as an ethical objector—right when the company was trying to secure an IPO or a massive ~$100 billion valuation, all while they quietly remained an active part of the war effort.
Beyond their military contracts, the podcast details several highly questionable business practices Anthropic has used to artificially inflate their numbers:
1. During a lawsuit regarding their military contract, Anthropic's CFO filed a sworn affidavit revealing the company had only made $5 billion in its entire lifetime. This directly contradicted leaked media reports suggesting they made $4.5 billion in 2025 alone. It revealed that the company's publicly perceived run rate was heavily exaggerated through the "shady revenue math" popular in Silicon Valley, a major discrepancy that most financial journalists ignored.
2. When the open-source agent library OpenClaw first launched, Anthropic deliberately allowed users to connect a $200/month "max account" and essentially burn through thousands of dollars of API compute at Anthropic's expense. Zitron points out that Anthropic knowingly let this happen to temporarily boost their usage metrics and hype while they raised a $30 billion funding round. Just weeks after securing the funding, they abruptly cut off access for these users, a move Zitron cites as proof of them being an "unethical company".
Furthermore, the company has faced criticism for gaslighting users, maintaining poor service availability, and silently degrading model performance while rug-pulling users on rate limits. As Zitron summarizes, it is highly unlikely that either Anthropic or OpenAI actually care about these ethical boundaries beyond how they can be weaponized for better PR and higher valuations.
There's some validity to these criticisms, but it would be a lot more credible to cite someone whose job isn't "loudly promote any claim that sounds negative for AI, regardless of how well-founded it is."
> Despite cultivating a reputation as the "ethical" AI company, Zitron argues that Anthropic's actions show they are just as ruthless and ethically questionable as their competitors.
Anthropic has taken 10s of billions from investors just like everyone else has. There is no such thing as "ethics" or "morality" when the scale of obligation is that large.
So yes, this is obvious despite whatever image they try to cultivate.
"LLMS are fundamentally incapable of controlling autonomous weapons" -- This was Anthropic's stance too, right?
"Quietly remained an active part of the war effort" - anthropic was totally transparent about it, but yeah not great.
"Leaks were wrong" - and that's Anthropic's fault?
OpenAI agreed to assist the DoD with zero boundaries and then lied about it. Can we at least give them credit for not doing that? If we just throw up our hands and say "they're all awful, whatever" then the result is reduced pressure on them to be better. Like it or not, I do not think AI is going away and as far as I can tell, despite billing problems, Anthropic's still the least bad frontier lab.
Did you even check the link? It's a podcast from Cal Newport, a quite known figure (at least in software engineering / compsci circles). So it's not exactly a random shitty podcast. And, it's also (obviously) not my content.
Agreed. they are better at the PR game. Some developers are grasping at straws looking for ways to not feel guilty and justify their usage of LLMs is from the "good guys". Anthropic is currently filling this role but eventually people will see behind the smoke and mirrors and release its not all that different from OpenAI or some of the other AI labs who are willing to sacrifice any amount of ethics if they mean they get the right paycheck or stroke their ego that they were on the team that built digital god.
I cancelled my subscription the minute they blocked access via OpenCode and switched to Ollama Cloud.
A bunch of people here tried to defend Anthropic, saying that it was justified because it was likely that Claude Code's harness had optimizations that would not be possible on OpenCode. It was clear from the source leak that nothing of this sort was the case, and that they were simply trying to avoid others distilling their models.
GLM and Queen are not on par with Opus, but they are good enough and I never had hit the usage limits, even with 2-3 sessions running.
they are essentially Lyft in early Uber vs. Lyft days. They are marketing themselves vaguely as being "better" because they're "more ethical" but their actions make it clear that they're not much better than OAI.
Except Lyft didn't kick you out in the bad part of town simply because you mentioned the word lollipop. Claude will terminate your session, peg you to 100% usage, and more, to stop you from using the service you paid for.
Ha. Yes. "Speedrunning enshittification" is the phrase that's been in my head.
The flat-rate plans were the top of the slippery slope to enshittification, really. If everyone were on metered billing there'd be no reason for all these opaque and sneaky attempts to limit usage. People would pay for what they get and get what they pay for.
There is nothing wrong with flat-rate plans. I work at an LLM-serving startup, and am aware of at least three competitors, that (a) provide flat rate subs (b) are extremely profitable and (c) are bootstrapped, ie. not beholden to investors (there are also many other competitors but I can't ascertain their profitability or investment status).
You simply need to price the flat-rate sub at a price that's profitable when averaged out over all of your users, both light and heavy, and prevent fully automated usage by the power users. That's it. This is immensely more user-friendly, and I doubt you'd get any traction at all if you didn't do this. Even if you pay more for the sub, having unlimited (non-automated) usage frees a mental barrier to using the product. If you have to pay for every request you make, it introduces a hesitation to do anything - it makes the user hesitant to experiment, hesitant to prompt for anything of slightly less significance, anxious about the exact token consumption of every prompt, and so on. It's not enjoyable to use when you're being penny pinched for every prompt.
Anthropic's problem, of course, is that they are not bootstrapped. They don't have a business model that can compete with startups running DeepSeek or GLM on their own hardware. Non-frontier startups got to skip the whole "tens of billions of dollars in debt" step of creating a frontier model from scratch, and still get to run a model that is perhaps 80%-85% as good as Anthropic's, which is good enough for millions of customers. So Anthropic is desperate, backed into a corner, and doing anything and everything they can to try to right their sinking ship, no matter how scummy.
Anthropic isn't backed into a corner. They have plenty of enterprise subscriptions. Individual user experience (especially billing) is suffering because it's not a priority in comparison. If they were as desperate as you described, they would try selling access to mythos.
The fact that they are adding code specifically to charge individual consumers more reeks of desperation. This isn't "individual users are suffering because they're lower priority and neglected", this is "individual users are being actively squeezed because Anthropic is desperate for every penny it can get".
This is such a stupid way to charge customers more. How many Claude code users use OpenClaw? Cheating customers is like burning down your house to keep warm. Anthropic aren't that stupid. I guarantee that this was some half-baked vibe-coded anti abuse system.
I do mind, since I enjoy speaking freely without concern of my opinions being linked to my employment. I assure you companies like this exist. Profiting off of inference is not the hard part, it's frontier training that is prohibitively expensive. You're free to disregard my commentary if you want, of course.
> Profiting off of inference is not the hard part, it's frontier training that is prohibitively expensive.
And given that Anthropic does both, it must make up its training costs by selling inference. jp57 was pretty clearly talking about Anthropic's flat-rate plans, rather than the flat-rate plans of companies that get to skip the most expensive part of the process.
I understand that very well, yes. The point I'm making is that I don't think Anthropic or OpenAI would have ever gotten significant traction if they didn't have flat-rate plans, because flat-rate plans themselves are not inherently predatory or part of the enshittification slope but actually extremely UX-friendly. Perhaps in another timeline, if their product was actually valuable enough to pay this price for, they could have simply provided a $50 plan as the standard level to provide enough margin to account for training costs as well. But as I see it DeepSeek is an existential threat to them, and they are now stuck between a rock and a hard place, because their product is devalued by its existence and if the frontier labs were to gate access with $50 plans they would get their lunch eaten even more quickly. It turns out there are downsides to burning inconceivably large stacks of other people's money.
> The point I'm making is that I don't think Anthropic or OpenAI would have ever gotten significant traction if they didn't have flat-rate plans...
That seems likely. If people had to pay their share of the actual all-in cost of the service (rather than having it be subsidized by investors with extremely deep pockets and a small handful of corporate customers), very, very few regular people would use it.
The point that 'jp57' pretty explicitly made [0] is that flat-rate plans that don't cover the all-in cost of providing the plans tend to result in those plans getting worse and worse and worse, as economic realities assert themselves. If the flat-rate plans that you are aware of actually cover the cost of providing the service, then you're discussing an entirely different situation that's entirely inapplicable to the discussion about Anthropic's pricing and degrading level of service.
[0] ...which is one that's understood by people who have been in pretty much any industry for more than a few years...
The crux of my argument is that there is a timeline where people would've paid the all-in cost of the service, with margin, as a flat-rate sub. The $20 rate was not sustainable when factoring in training costs but if not for DeepSeek they could have simply raised the prices rather than gestures broadly whatever the fuck is going on at Anthropic now, with a new PR fumble every three days. If the Chinese models didn't exist, people would've groaned but would likely still pay $40 or $50 for an LLM subscription.
You misdirected my quoted statement to assert a position I did not take. When I talk about flat-rate subs being a good UX, I am not talking about at a subsidized rate. My position is that people will pay more for a flat-rate sub than they are willing to through per-token billing. That is, a consumer who would only pay average $10/mo if they used the API will voluntarily pay $20/mo for a sub, because even though it's a worse value the latter is a tremendously more friendly user experience. When I say that flat-rate subs are necessary for traction, I mean that solely from a user experience perspective, not "subsidized usage is necessary for traction".
I'm stepping away from LLMs in general and did cancel Claude code subscription this month because I respect myself very much and I deserve a better and transparent treatment.
If you must - in my experience Deepseek v4 is incredible value in every aspect. Pricing is transparent.
But like I said, I have funds in different AI gateways but I'm preferring to write by hand because I don't want surprising bugs and unnecessary code in my end result.
I’ve got a NixOS Qemu VM I use to run openclaw in. I had Claude help me set it up, and it runs local models on my own machine in a config based sandbox.
Why should Claude block or charge extra to work on that?
Why should Claude care if I have instructions for Hermes or OpenClaw in my project repos?
This fingerprinting is incredibly sloppy for how much access to a machine Claude code has.
Now you've learned the advantage of knowing how to do things yourself. When you depend on untrustworthy agents, you shackle yourself to their idiotic whims. Be careful who you partner with.
I run an OpenClaw VM and used Claude Code to build the VM scripts. The VM is connected to local llama.cpp, so OpenClaw and the models are running on my own physical hardware.
That is a huge red-flag. While I understand that they will do some policing/censoring, this is way beyond what I would consider acceptable.
They can have a different price plan for agentic stuff, but these things where they “accidentally” whoops match on specific keywords and trigger extra usage charges is giving a evil-microsoft-vibe
They are trying to make a moat where no possibility of creating a moat exists.
It’s a huge mistake at the level of IBM trying to reestablish dominance over PCs by making MicroChannel the new standard; this failed horribly and cost IBM its market leadership and reputation.
MCA was technically better at the time, but the industry responded with EISA and VLBus which led to PCI and today’s PCIe.
Things like these (Google also banned me from Antigravity for briefly using an agent) and the massive quality swings made me cancel all 3 subs last week and resort to my local Qwen 3.6 only. Open models are already great and only getting better, and I really enjoy the privacy and consistency of a model I run myself.
I don't think anyone is questioning all the benefits of using local LLMs. Those are readily apparent.
I just don't believe for an instant that they're anywhere in the same ballpark of capabilities as running Opus or similar. My time is the most valuable resource. Opus would need to be SIGNIFICANTLY more costly and unstable for me to start entertaining local models for day-to-day development.
Perhaps whatever work you're doing makes this trade-off more sensible, but I struggle to see how that could be true. I'm averse to running Sonnet on a large amount of software engineering problems - let alone Qwen.
I think you'd be surprised, I find that the harness is what makes the real difference. I also prefer to be on the loop, actively guide and review. Local models are definitely much less autonomous as of today so if you need to be churning out code at speed they're probably not for you.
If you know what you're doing and prompt it correctly, local models are great. If you're just vibe coding and relying on the LLM to fill in all the gaps for you and basically build the software for you, yeah you need SOTA to deal with that.
For now we infer through few weights, lossily; but then in full precision. Now I represent in part; but then shall I represent as fully as the data was sampled.
I have a 64GB M1 Ultra dedicated to llama.cpp. I get 40 tok/s on a fresh session, decreasing slowly to about 25 tok/s at around 50% of the 256K context, then down to 20 tok/s or less beyond that, but I rarely let it go much higher and handoff instead. This is whith Qwen 36B A3B at 8Q without KV quantization. It's not super fast but perfectly usable for me.
Subscription models only work when marginal costs are low and/or there’s a good variety of usage that roughly averages out. Or, you need to be able to kick out abuse.
Unfortunately for those of us who just want to eat a nice filling meal at the fixed price all you can eat buffet of AI subscriptions, a minority of customers keeps paying for the all you can eat buffet and staying for hours and bringing containers to sneak food out when they leave. And they keep wearing disguises to try and evade detection.
It’s a losing battle for the provider, which ultimately means the subscription pricing model can’t work, which hurts the majority of customers that just want to use the system as intended and no longer have a subscription model available.
I have plenty of frustrations with Anthropic as a paying customer, but tbh is doesn’t strike me as all that awful, just some annoying collateral damage. I’d rather have that than no subscription model at all.
I wouldn't be surprised if the AI usage model moves towards a bidder/auction model. Set how much you'd willing to pay for your AI request, and they evaluate requests starting from the highest to lowest bids.
It definitely would make sense, especially if they are capacity constrained, but it’s also a losing PR move for whoever moves first in the space unless the big players all shift at the same time.
Who remembers the Google of Eric Schmidt and "Don't Be Evil"?
The truth is that it doesn't matter what companies say, what they claim, what they do, and what their CEO says/claims/does.
It's just a matter of time until the shareholders will get the right CEO to maximize shareholder value.
People in the comments who want a statement or a "reorientation" or a commitment from Anthropic leadership are missing the principles of how capitalism functions. Shareholder value cannot be compromised. In every battle between morality and profit, values and profit, public good and profit, ultimately all things will mutate into a state to allow profit will prevail. Always.
I saw a talk by Boris where he said, basically that Claude codes itself now. They have it automatically writing features and reviewing PRs, apparently. I suspect that much of the code has never been seen by human eyes within Anthropic.
their CEO has been shouting from the rooftops that programming is dead. ofc that would ripple down the org chart and result in a culture of bad programming.
they're just holding it wrong.. what model are they using? they should make sure they're on Opus 4.5+. That was a stepwise improvement and was when AI coding clearly became the futureₖₑₖ
why do people want to continue to use anthropic despite their shitty service? its not like they have some kind of lock-in as it is still new company and it has shown its color before we are stuck with it unlike google/meta etc.
Totally agree. This is why open source models and toolings are so important for the ecosystem. I would not want these companies decide what we can or cannot do.
When compute poverty hits these big labs it’s all going to be the same. The ping pong tables and drinks fridges disappear.
The only thing they can hope for is to maintain momentum and critical mass long enough to find ways to pay for all this or have Moores law make the average user request become economical.
possibly related, it errors if my working directory is a checkout of OpenCode. i was using CC to work on some patches for OC and had to work in a parent directory and then tell Claude to work on the files inside the "opencode" folder.
I think it's obvious that they are critically lacking in compute capacity especially since OpenAI has committed billions to locking up all the future compute production.
And I don't necessarily think it's wrong for Anthropic to introduce QoS or throttling on users of their models. It's pretty much a necessity when offering public access to a scarce resource and it's been a common practice for decades.
What is the alternative? We just accept that it doesn't work half the time because the system is overloaded with molt bots?
I agree. If compute is the issue and pricing can't budge then something has to give.
They would have kept my business if they were honest and upfront. Instead they sold me something that worked well, broke it without warning, remained silent about it until enough people caught on, chose to do nothing, then proceeded to release a model that eats ~30% more tokens with no advantage over prior models.
If they chose to unbrick their model and offered what we had a couple months ago at a 50% hike, I would have been onboard. I've seen enough now of how this company treats its customers to continue using or recommending them.
Also, Codex works much better than CC now for anyone who happens to be on the fence.
The alternative is to price their product transparently. If there is too much demand and supply is limited: Charge more.
Anthropic wants to have their lunch (low apparent prices, increased market share) and eat it too (controlled costs, adequate production to serve the demand).
They're advertising themselves as a $5 All-You-Can-Eat buffet, but then aggressively and arbitrarily restricting admission, sneakily swapping out the high-quality ingredients for garbage-tier slop, and kicking out anyone who even utters the words "to go box" or "doggie bag".
I cancelled my subscription so not really defending them myself but if all of their customers were humans who used it normally I bet they could serve everyone. It's when someone presses a few keys walks away and a bot uses tokens for 72 hours straight that it becomes a problem. Then people buy 3 accounts and do that for weeks at a time.
Could you do that as a human? Sure but you'd likely burn out after a couple of weeks. Also the human would probably use those tokens far more effectively and would not need as many. It's feels the same as someone installing a crypto miner on their servers in my mind. Abhorrent behavior.
I find it incredibly that after all the good faith Claude Code built during 2025 they are destroying users trust is such amateurish ways (same as hermes.md)
You say that like it's a gotcha. I think the fact that they reached 2B/mo in revenue by dogfooding cc is all the proof that one needs that this thing actually works. In fact it works so well that more people want it than they can serve. For months now they've been having issues when EU and US tz are both online at the same time.
> I think the fact that they reached 2B/mo in revenue by dogfooding cc is all the proof that one needs that this thing actually works.
That's a notable achievement, but let's have some balance... It's also responsible for the biggest self-own in software industry history by leaking their 1) crown jewels (i.e., source code) 2) the existence of their next model Mythos, and 3) their roadmap in a highly competitive market.
Eh... I personally think that having the keypads to enter a DC running on DNS served by that same DC is a bit more self-owning than leaking the source code of an app, but I get your point. It's obviously not perfect, but it's also obviously working.
Let's put this in perspective. Imagine it's 3 years ago, April 2023. Chatgpt has been launched for 4 months. We've all been using it, and writing poems in parrot talk or whatever. Someone tells you "In 2 years time there will be an app that lets you use LLMs to write code. It will be coded by humans for 3 weeks, then by humans + LLMs for 6 months, and then by LLMs mostly unsupervised. One year after that, they'll be making 2B/mo out of that app". Would you believe them? Not even the most maximalist, overhypers, AI singularity frenzied crazy people would have said that. And yet... it happened.
Is the reason they reached 2B/mo partially contributed by the fact that their users feel like they get unlimited use of it?
If ‘feeling like it is unlimited use’ is a huge part that creates the 2B/mo, this change of limit might jeopardize it.
That being said, Anthropic can be diverting capacity to train the next model, and if it is significantly better, people would start flocking back again.
The problem with slop is, nobody understands it. Nobody ever designed it, nobody really knows how it works. You’re just putting blind faith in the slop you’ve shipped.
It lets you be very quick, but if you’ve accidentally compromised all your data or bank accounts through the slop then you won’t know until you’re destroyed.
Claude is bad for business....that is painfully obvious.
At this point I assume you are coping with having drank the koolaid and fired key staff believing claude will replace them...back when it was cheap....because nearsightedness affects decision makers much more during hype cycles......
Having had Claude Code jump to inserting juvenile and all-filtering regex to (attempt to) solve open-ended semantic natural language problems (-sigh- there's 12 hours of my life I'll never get back), I can absolutely imagine that this was someone trying to code up a "defense in depth" mechanic that was explosively insufficient after Claude Code (even Opus 4.6) made a series of faulty assumptions.
This one feels like prime space for Hanlon's razor: "Never attribute to malice that which is adequately explained by stupidity."
The hassle with the performance of these systems is that they're ~70% of the way to awesome. For advanced prototyping (my current job description), a fast 60% of awesome is groundbreaking and game-changing. For production and real businesses, that last 30% is a really, really important thing to figure out.
Anthropic is losing a ton of goodwill by not being more honest about their constraints. They've been buckling under load for months, and instead of doing the most honest thing (keep weekly usage limits same, make 5 hour usage limits have surge pricing where the usage-cost of X tokens is scaled based on dynamic load), they're doing a lot of hacky things to try to get a similar effect. I suspect they feel the optics of being honest would be too bad, so instead it's a slow bleed where they piss off users one by one
Ok I am usually defending Anthropic, but it seems like this OpenClaw and Hermes ban was implemented incredibly poorly; it looks like a simple regex.
Didn’t they think about “we need to make sure Claude Code is never banned” ? Could have been as easy as including some Claude Code specific prompting traits (tools, system prompt, whatever) in there and automatically whitelisting it.
Is it foolproof? No. Will it avoid banning legit users? Absolutely.
First do the first large sweep, then see what still falls through, then ban those.
It really seems they were panicking due to capacity and there was very little oversight with all this.
Oh come on Anthropic, just admit straight away that any other pricing than usage-based is completely unsustainable and is being phased out.. maybe doing it once but officially could save you some brand damage.
what a company with really bad customer practices. I'm really glad I moved entirely to open source models. if you're disgusted by these practices as I am, I really recommend you use opencode (or any of the other 20 agents) and the GLM 5.1, or Kimi K2.6 or Deepseek V4 Pro models. You will be shocked how effective they are.
haven't used claude in about 2 weeks and I do not miss it.
Interesting people talking about whether they should be "defended," here or whatnot, and all of that strikes me as wildly naive.
They have a business model that's more or less known, and that includes THEIR AI model(s) that they get to put out there however they want. I don't like it much at all, I actually sort of like the idea that they "owe" more because they probably "stole" a bunch of stuff to get the thing going.
But I mean, don't be mad, be proactive. Anthropic is going to try to Microsoft this in whatever way possible, and we all see that the numbers don't really add up.
Asking them pretty please to be nicer, meh. Let's figure out better, and more free-software-like ways to do this.
I think that’s an ok move, definitely better than canceling code on pro users for example, I would support to even have a new pricing tier only for openclaw, so they don’t ruin the usage on others. I noticed the ones who use claude code usually are software developers or sysadmins, meanwhile most openclaw ones are your average HR stacy and lazy middle managers, so yeah, it should be a separate tier for them.
openClaw does so muhc more then Claude code tbh, running 9 agents from the one machine, schedual some tasks, add some personal personas for each agent, claudeCode (which i like alot) is on rails, openClaw is full openworld.
I reproduced this on my account.
Immediate disconnect and session usage went to 100%That's malicious and I think this is scamming from the literal money (you didn't do anything wrong, you executed one command and they scammed you out of the fair usage you paid for).
Please raise the ticket or at least GitHub issue for visibility.
Sooner or later some sort of complaint to the relevant trade authority should happen - this is a scam operation at this point.
There are many possible explanations for this outcome to have occurred other than malice. If you're an engineer by trade, consider how many bugs you've been responsible for over the course of your career that you didn't intend. Probably a lot.
How about we turn down the heat, everyone?
There's been a sustained pattern of incidents. If Anthropic were truly serious about not wanting to take people's money, then they would have put in place whatever review processes were necessary to stop this from happening. So regardless of whether or not they specifically intend to cause harm, they're willingly letting it happen, which is just about as bad.
Yes, it's reasonable to turn down the heat. But it's also reasonable for people to be upset when their money is taken from them, and when the company that does so is effectively beyond persecution for doing so.
> How about we turn down the heat, everyone?
How about Anthropic turn down the heat and refunds money to everyone for every bug it created with its LLM?
Even with the best of faiths, this is at the very least a shoddily vibe coded “detect and low-key block attempts to use Claude for Openclaw” - it decided to look for specific strings wrapped in json without realizing this doesn’t always imply it’s an actual payload for Openclaw itself. And the human driving it was too dumb to review/catch this bad inplementation.
So maybe not malice, but certainly a level of ineptitude I don’t expect from a crucial vendor from a tool that’s become essential for many developers.
(I don’t care, I do just fine when Claude is down or refuses to help me (it has happened) though)
And the stealing of $200 here? More non malice?
https://github.com/anthropics/claude-code/issues/53262#issue...
Yeah they probably just typed in "Hey Claude, figure out a way to get our inference spend under control - no mistakes!" and shipped it
Also they ain't wrong. In what other context does OpenClaw get mentioned?
"You may not use our service if you mention OpenClaw" is a harsh line but hardly illegal or forbidden any more than any other service restriction (i.e. no use allowed for high-stakes financial modeling). Don't like it, cancel your plan.
> is a harsh line
But that's the thing -- there is no line! Where is this specified? How can we know what service restrictions there are? For all I know, my plan could be exhausted at any point during the workday just because I happened to touch on some keyword Anthropic has decided to ban.
> Don't like it, cancel your plan.
Ah, but I thought these models were supposed to have been trained for the sake of humanity? That the arbitrary enclosure of the collective intelligence was for our own good? These concepts are not compatible.
> How about we turn down the heat, everyone?
The heat is coming, in part, from the lack of a proper support channel.
Nuance? Ignorance vs malice? You think too highly of folks.
Its not Claude Code.
Its "Fraud Code".
All of this is just criminal and fraudulent behavior, done July a whole bunch of people who haven't learned their lesson, and keep sending Anthropic more money for abuse at scale.
It's in the TOS, so no, not fraud. You might not like it that Anthropic doesn't want you running OpenClaw (effectively owned by a competitor) on CC, but that doesn't make it fraudulent or criminal.
There is literally nothing close to illegal about this behavior. You read the terms of service right, which provides a long list of explicit and implicit disclaimers?
So, in America, just because it's written in a contract does not mean it's enforceable in anyway.
I can make you sign a infinitely generating contract, that doesn't mean it's enforceable/
Think they turned it off, or it's not always active. I can't reproduce it myself.
Not reproing here either.
Or a/b testing.
That's rather shitty. It's one thing to disallow bypassing preferential pricing models, it's a completely different thing to castrate your model against some uses.
You can see how it goes in the future. Wanna vibe code a throwaway script? $0.20. Ah, it's for a legal document search? $10k then. Oh and we'll charge 20% of your app sales too - I can see how they are going in real time, mind you!
Unironically yes.
I predict that costs will grow to 80% of what it would cost a human, across the board for everything AI can do.
"It's still cheaper than a human" they'll say. Loudly here on HN too.
Of course this will happen slowly, very slowly. Lets meet again in 10-20 years.
> "It's still cheaper than a human" they'll say.
The question is how much friction there will be for people to switch over to Gemini, GPT or maybe even DeepSeek or Mistral or whatever. Even if price hikes are inevitable across the board, the moat any single org has is somewhat limited, so prices definitely will be a factor they'll compete on with one another at least a bit.
This is assuming there will be no competition. But why wouldn't there be? Especially since you can use open source models, which are not too far from frontier models (from now).
Its not20 years. Its now. Nvidia has already said that tokens cost more than humans.
https://finance.yahoo.com/sectors/technology/articles/cost-c...
So like taxes except they actually help you survive?
I'm not a lawyer but is this legal? It's extremely anticompetitive.
what is illegal about it?! their product, they can do whatever they want and you can choose to be a customer or not, no?
They are technically billing people for services not rendered without any disclaimer?
Price discrimination for services is mostly legal
This is absolutely how it’s going work. AI loses way too much money to not be enshittified.
It’s a way less transformational technology when put in context of the real price tag.
No chance unless open weight models out of China discontinue. The gap right now is practically nonexistent.
When the consolidation phase starts, you bet your ass open weight models are going to stop.
I don't think consolidation will ever happen, the AI space is already dominated by a few whales.
Seems most of the open weight models are from outside the USA (shocker), going to be interesting to see how THAT shakes out.
Deepseek has demonstrated that there is no reason for it to actually lose money. The awful business practices and monopoly tactics of the frontier model labs in the US are the problem.
I mean obviously. Why would the companies that control this technology NOT charge the absolute maximum amount their customers are willing to pay?
This doesn't even have anything to do with if it loses money or not. Obviously they are going to charge as much as possible.
I think it goes beyond this. I was just using claude to edit a blog post which mentioned OpenClaw and I got this response: "The "OpenClaw" reference — I assume that's a typo or playful reference; if you mean a real product, I couldn't find it under that spelling and you'll want to fix or footnote it.". I gave it a direct link to openclaw.ai and the chat instantly ended and hit my 5hr usage limit. Could have been a coincidence, but I had only lightly been using sonnet in the morning so it seems unlikely. Very odd.
Dragons steal gold and jewels... and they guard their plunder as long as they live... and never enjoy a brass ring of it. Indeed they hardly know a good bit of work from a bad, though they usually have a good notion of the market value
My theory is the dragons actually benefit immensely from sitting atop the gold piles as it acts as an amazing heat sink.
I don’t think that really fits with the metaphor but I wanted to say my piece regardless.
We don’t really have dwarven gold hoards anymore - I’m thinking we can prove climate change is caused by overheating dragons.
Everyone send me all your gold and I’ll prove it.
It doesn't look like anything to me
The weird thing is that it found sources for all of my other claims and references no problem, but acted like it didn't know what openclaw was when openclaw.ai is the first thing that pops up on google.
Same past days it sometimes tried to gaslight me saying OpenClaw isn't a thing.
Lmao, I can 100% believe that they are deliberately filling your usage bar to sabotage their competition. These people have no morals.
"Sorry, that was a bug!" Thariq will be on scene shortly, don't worry.
Yeah it will be something like "we A/B tested on 0,05% of users and ..."
I mean that also just sounds illegal...
It also sounds extremely counterproductive to try and sabotage your competition by.. driving your customers away? I have no love for these companies but it's a silly conclusion to jump to.
They don't want customers that make them bleed more money than they're supposed to.
Not if a chatbot did it, maybe. No legal precedence here. Also they are a defense and offense contractor they could kill people and nothing would happen
Would they act differently if it was?
Claude.ai is now at a 98.85% uptime. There's been so many frustrations with Claude / Anthropic lately (very heavy usage limits, wrong A / B testing, etc.).
Claude status: https://status.claude.com/
I have been really happy with my Codex subscription lately, but feels like these things change every other day. The OpenCode Go subscription for trying out GLM, Kimi, Qwen, Deepseek and friends also looks useful.
But nonetheless, Opus 4.6 is a very capable model, but justifying a Claude subscription gets more and more difficult, think I might just sometimes use it through OpenRouter or as part of something like Cursor (although I'm not sure about the value of that subscription as well).
OpenCode Go: https://opencode.ai/go
Cursor: https://cursor.com
There were periods where I was entirely unable to use Claude Code for hour+ due to auth gateway always returning 500 or timing out, there was an "elevated errors" incident shown on status.claude.com, but zero minute of downtime recorded (not even "partial outage"). So the real uptime should be even worse.
April has been a crazy month for open weights models. I've been using Claude Code for work and Kimi 2.6 for personal projects and Kimi has been very good. Glm-5.1 is also great. Qwen, Mimo and Deepseek I need to test some more, but they all have been producing good results. I have the impression that they are all are at the same level, or close to, Sonnet 4.6.
What are you running them on?
They are close to Opus, not Sonnet.
The little qwen36 is at sonnet level . Kimi2.6 is about opus. The one can run on a single GPU on your gaming pc. The other you can run way cheaper from a provider. Or if you are really wealthy and have lots of gpus can run it yourself.
Not sure where deepseek 4 sits
Yes multiple GPUs absolutely help with inference even for a single model instance. Some models are simply too big to fit on the largest available GPU.
Check out tensor parallelism
Would "lots of gpus" even help for huge models? Maybe this is exposing my lack of knowledge but don't you need to keep the whole model and context in a single GPU's VRAM? My understanding is that multiple GPUs help with scaling (can handle N X inference requests simultaneously) but it doesn't help with using large models. If that were the case, I could jam another GPU in my box and double the size of model I can serve.
> Would "lots of gpus" even help for huge models? Maybe this is exposing my lack of knowledge but don't you need to keep the whole model and context in a single GPU's VRAM?
How do you think the large providers do inference? No single GPU has 1TB plus of memory on board. It’s a cluster of a bunch of gpus.
Please don't oversell them. Eg Kimi k2.6 has a maximum context size of 270k, that's a quarter of opus.
The model is fine, Ive switched to it entirely for a personal project, but it's not opus.
And no, you're not running then locally unless you're a millionaire. You still need hundreds of GB (500+++) of VRAM on your graphics card - that's not at a level of consumer electronics.
Sure you can run the quantized models, but then you're at Haiku performance.
Qwen 3.6 runs in a single GPU. But I mostly agree with you
> Claude.ai is now at a 98.85% uptime.
So, at least better than GitHub, right? :)
This is very concerning. Their heavy handed tactics haven't impacted me personally yet but I am increasingly nervous and casting about for viable egress paths if I need to flee Claude Code. I really hope they pump the breaks and thoroughly reorient themselves. They are under a lot of competing pressures and probably can't make a decision that won't upset a lot of people (in order to balance growth and capacity etc), but are coming to the worst possible conclusions.
For instance, maybe you can't afford to take on more customers right now, Anthropic. Maybe if you are severely undermining the customer relationships you already have, you should just admit you can't sell any more 20x plans right now and only accept new customers at lower tiers until you have the necessary capacity.
This is also a DoS you could drive a truck through, and it's disturbing such an obvious vulnerability was shipped at all.
> or instance, maybe you can't afford to take on more customers right now, Anthropic. Maybe if you are severely undermining the customer relationships you already have, you should just admit you can't sell any more 20x plans right now and only accept new customers at lower tiers until you have the necessary capacity.
Or just increase prices for new claude code users? Surely transparent upfront across the board price increases are easier to swallow than hidden context-based pricing changes like this?
> casting about for viable egress paths if I need to flee Claude Code
Check out OpenCode (the OSS product [1]) and OpenCode Go/Zen (the LLMaaS [2]). Use a more expensive model with larger context (like GLM-5.1) for orchestration and cheaper models for coding and iteration on acceptance criteria (writing and passing tests). I also throw a more expensive vision-capable model into the mix like Gemini 3 Flash to iterate on UI tasks using Playwright. With the base usage in Go and pay as you go on cheaper models like MiniMax you can get a lot done for not a lot of coin.
[1] https://github.com/anomalyco/opencode
[2] https://opencode.ai/go
Codex has been great for me
Anything coming from OpenAI is an automatic "Hell, no!" for me.
I really want to stick with A\ given everything known about Altman, but man are they speedrunning the "how to destroy your reputation" guidebook.
They have better PR than OpenAI but they are not a more ethical company. They do a bunch of shady stuff and are just as much involved in military applications. Cal Newport’s recent podcast had a good discussion about this: https://youtu.be/BRr3pAPsQAk?si=jaRJYJ_XQE7VpxPN
Pet peeve of mine is people saying "hey this thing is totally shady/false, I've got proof right here <links to hour long podcast>".
It happens surprisingly often.
I understand not everyone has the interest or time to sit through an hour long podcast. But last I checked this is HN, and I think that podcast is right up the alley for many of us here. Cal Newport is not exactly a 'random podcaster'.
Next time I can summarize some of the talking points in my comment though, but I didn't want to poorly regurgitate the arguments when they were readily available in the video lol.
Although I see another poster has commented the key takeaways :)
Cal Newport and tech commentator Ed Zitron discussed this disparity between Anthropic's public image and their actual practices. Despite cultivating a reputation as the "ethical" AI company, Zitron argues that Anthropic's actions show they are just as ruthless and ethically questionable as their competitors.
Anthropic has been deeply integrated with the US military, having been installed with classified access since June 2024. The podcast highlights that Claude has been actively utilized during the "Venezuela incursion" and the ongoing "war in Iran".
Despite this active involvement, CEO Dario Amodei released a statement attempting to publicly distance the company from the Department of Defense by declaring they would not allow their technology to be used for "mass domestic surveillance" or "fully autonomous weapons". Zitron categorizes this as a highly calculated PR maneuver, pointing out that LLMs are fundamentally incapable of controlling autonomous weapons anyway. The stunt successfully manufactured a wave of positive press—with celebrities and commentators praising Anthropic as an ethical objector—right when the company was trying to secure an IPO or a massive ~$100 billion valuation, all while they quietly remained an active part of the war effort.
Beyond their military contracts, the podcast details several highly questionable business practices Anthropic has used to artificially inflate their numbers:
1. During a lawsuit regarding their military contract, Anthropic's CFO filed a sworn affidavit revealing the company had only made $5 billion in its entire lifetime. This directly contradicted leaked media reports suggesting they made $4.5 billion in 2025 alone. It revealed that the company's publicly perceived run rate was heavily exaggerated through the "shady revenue math" popular in Silicon Valley, a major discrepancy that most financial journalists ignored.
2. When the open-source agent library OpenClaw first launched, Anthropic deliberately allowed users to connect a $200/month "max account" and essentially burn through thousands of dollars of API compute at Anthropic's expense. Zitron points out that Anthropic knowingly let this happen to temporarily boost their usage metrics and hype while they raised a $30 billion funding round. Just weeks after securing the funding, they abruptly cut off access for these users, a move Zitron cites as proof of them being an "unethical company".
Furthermore, the company has faced criticism for gaslighting users, maintaining poor service availability, and silently degrading model performance while rug-pulling users on rate limits. As Zitron summarizes, it is highly unlikely that either Anthropic or OpenAI actually care about these ethical boundaries beyond how they can be weaponized for better PR and higher valuations.
There's some validity to these criticisms, but it would be a lot more credible to cite someone whose job isn't "loudly promote any claim that sounds negative for AI, regardless of how well-founded it is."
[delayed]
> Despite cultivating a reputation as the "ethical" AI company, Zitron argues that Anthropic's actions show they are just as ruthless and ethically questionable as their competitors.
Anthropic has taken 10s of billions from investors just like everyone else has. There is no such thing as "ethics" or "morality" when the scale of obligation is that large.
So yes, this is obvious despite whatever image they try to cultivate.
Anthropic is a public benefit corporation which limits liability to shareholders.
Just because they screwed up their billing doesn't mean every ethical commitment they've ever made is bunk.
> There is no such thing as "ethics" or "morality" when the scale of obligation is that large.
At that scale, ethics and morality should become more important, not discarded
Alternatively, finance at that scale ought not be permitted to exist, because of the moral hazard it represents.
You will find that morals and ethics at that scale are too expensive to maintain.
"LLMS are fundamentally incapable of controlling autonomous weapons" -- This was Anthropic's stance too, right?
"Quietly remained an active part of the war effort" - anthropic was totally transparent about it, but yeah not great.
"Leaks were wrong" - and that's Anthropic's fault?
OpenAI agreed to assist the DoD with zero boundaries and then lied about it. Can we at least give them credit for not doing that? If we just throw up our hands and say "they're all awful, whatever" then the result is reduced pressure on them to be better. Like it or not, I do not think AI is going away and as far as I can tell, despite billing problems, Anthropic's still the least bad frontier lab.
I think all the AI companies want to hook up with the US military, as it's the only way they'll cover their debt. For investors.
"You must destroy the economy to keep us afloat, because National Security!" has been a clear goal of the LLM hucksters for a long time.
Probably some Slopcoded bot which posts fake comments to drive people to their content.
After all, if you’re paying hundreds of millions to buy these shitty podcasts, you might as well host some bots.
Account is from 2016 with 6k karma? : doubt:
Did you even check the link? It's a podcast from Cal Newport, a quite known figure (at least in software engineering / compsci circles). So it's not exactly a random shitty podcast. And, it's also (obviously) not my content.
Agreed. they are better at the PR game. Some developers are grasping at straws looking for ways to not feel guilty and justify their usage of LLMs is from the "good guys". Anthropic is currently filling this role but eventually people will see behind the smoke and mirrors and release its not all that different from OpenAI or some of the other AI labs who are willing to sacrifice any amount of ethics if they mean they get the right paycheck or stroke their ego that they were on the team that built digital god.
I cancelled my subscription the minute they blocked access via OpenCode and switched to Ollama Cloud.
A bunch of people here tried to defend Anthropic, saying that it was justified because it was likely that Claude Code's harness had optimizations that would not be possible on OpenCode. It was clear from the source leak that nothing of this sort was the case, and that they were simply trying to avoid others distilling their models.
GLM and Queen are not on par with Opus, but they are good enough and I never had hit the usage limits, even with 2-3 sessions running.
What's just as crazy is people defending ollama.
https://github.com/ollama/ollama/issues/3185
they are essentially Lyft in early Uber vs. Lyft days. They are marketing themselves vaguely as being "better" because they're "more ethical" but their actions make it clear that they're not much better than OAI.
Except Lyft didn't kick you out in the bad part of town simply because you mentioned the word lollipop. Claude will terminate your session, peg you to 100% usage, and more, to stop you from using the service you paid for.
Ha. Yes. "Speedrunning enshittification" is the phrase that's been in my head.
The flat-rate plans were the top of the slippery slope to enshittification, really. If everyone were on metered billing there'd be no reason for all these opaque and sneaky attempts to limit usage. People would pay for what they get and get what they pay for.
There is nothing wrong with flat-rate plans. I work at an LLM-serving startup, and am aware of at least three competitors, that (a) provide flat rate subs (b) are extremely profitable and (c) are bootstrapped, ie. not beholden to investors (there are also many other competitors but I can't ascertain their profitability or investment status).
You simply need to price the flat-rate sub at a price that's profitable when averaged out over all of your users, both light and heavy, and prevent fully automated usage by the power users. That's it. This is immensely more user-friendly, and I doubt you'd get any traction at all if you didn't do this. Even if you pay more for the sub, having unlimited (non-automated) usage frees a mental barrier to using the product. If you have to pay for every request you make, it introduces a hesitation to do anything - it makes the user hesitant to experiment, hesitant to prompt for anything of slightly less significance, anxious about the exact token consumption of every prompt, and so on. It's not enjoyable to use when you're being penny pinched for every prompt.
Anthropic's problem, of course, is that they are not bootstrapped. They don't have a business model that can compete with startups running DeepSeek or GLM on their own hardware. Non-frontier startups got to skip the whole "tens of billions of dollars in debt" step of creating a frontier model from scratch, and still get to run a model that is perhaps 80%-85% as good as Anthropic's, which is good enough for millions of customers. So Anthropic is desperate, backed into a corner, and doing anything and everything they can to try to right their sinking ship, no matter how scummy.
> prevent fully automated usage by the power users.
But being a power user and fully automating things is the whole appeal.
Anthropic isn't backed into a corner. They have plenty of enterprise subscriptions. Individual user experience (especially billing) is suffering because it's not a priority in comparison. If they were as desperate as you described, they would try selling access to mythos.
The fact that they are adding code specifically to charge individual consumers more reeks of desperation. This isn't "individual users are suffering because they're lower priority and neglected", this is "individual users are being actively squeezed because Anthropic is desperate for every penny it can get".
This is such a stupid way to charge customers more. How many Claude code users use OpenClaw? Cheating customers is like burning down your house to keep warm. Anthropic aren't that stupid. I guarantee that this was some half-baked vibe-coded anti abuse system.
> prevent fully automated usage by the power users
this is a non-starter
I also assume that forcing usage to spread out, via those 5-hour windows, has cost advantages.
LLM serving startup => bootstrapped => extremely profitable
Mind sharing a link?
I do mind, since I enjoy speaking freely without concern of my opinions being linked to my employment. I assure you companies like this exist. Profiting off of inference is not the hard part, it's frontier training that is prohibitively expensive. You're free to disregard my commentary if you want, of course.
Why not just name one of those three competitors?
> Profiting off of inference is not the hard part, it's frontier training that is prohibitively expensive.
And given that Anthropic does both, it must make up its training costs by selling inference. jp57 was pretty clearly talking about Anthropic's flat-rate plans, rather than the flat-rate plans of companies that get to skip the most expensive part of the process.
I understand that very well, yes. The point I'm making is that I don't think Anthropic or OpenAI would have ever gotten significant traction if they didn't have flat-rate plans, because flat-rate plans themselves are not inherently predatory or part of the enshittification slope but actually extremely UX-friendly. Perhaps in another timeline, if their product was actually valuable enough to pay this price for, they could have simply provided a $50 plan as the standard level to provide enough margin to account for training costs as well. But as I see it DeepSeek is an existential threat to them, and they are now stuck between a rock and a hard place, because their product is devalued by its existence and if the frontier labs were to gate access with $50 plans they would get their lunch eaten even more quickly. It turns out there are downsides to burning inconceivably large stacks of other people's money.
> The point I'm making is that I don't think Anthropic or OpenAI would have ever gotten significant traction if they didn't have flat-rate plans...
That seems likely. If people had to pay their share of the actual all-in cost of the service (rather than having it be subsidized by investors with extremely deep pockets and a small handful of corporate customers), very, very few regular people would use it.
The point that 'jp57' pretty explicitly made [0] is that flat-rate plans that don't cover the all-in cost of providing the plans tend to result in those plans getting worse and worse and worse, as economic realities assert themselves. If the flat-rate plans that you are aware of actually cover the cost of providing the service, then you're discussing an entirely different situation that's entirely inapplicable to the discussion about Anthropic's pricing and degrading level of service.
[0] ...which is one that's understood by people who have been in pretty much any industry for more than a few years...
The crux of my argument is that there is a timeline where people would've paid the all-in cost of the service, with margin, as a flat-rate sub. The $20 rate was not sustainable when factoring in training costs but if not for DeepSeek they could have simply raised the prices rather than gestures broadly whatever the fuck is going on at Anthropic now, with a new PR fumble every three days. If the Chinese models didn't exist, people would've groaned but would likely still pay $40 or $50 for an LLM subscription.
You misdirected my quoted statement to assert a position I did not take. When I talk about flat-rate subs being a good UX, I am not talking about at a subsidized rate. My position is that people will pay more for a flat-rate sub than they are willing to through per-token billing. That is, a consumer who would only pay average $10/mo if they used the API will voluntarily pay $20/mo for a sub, because even though it's a worse value the latter is a tremendously more friendly user experience. When I say that flat-rate subs are necessary for traction, I mean that solely from a user experience perspective, not "subsidized usage is necessary for traction".
Adding many new chapters to it
I think people inside the tech bubble don't realize that AI companies are considered villainous by the public. So there's no reputation to destroy.
I'm stepping away from LLMs in general and did cancel Claude code subscription this month because I respect myself very much and I deserve a better and transparent treatment.
If you must - in my experience Deepseek v4 is incredible value in every aspect. Pricing is transparent.
But like I said, I have funds in different AI gateways but I'm preferring to write by hand because I don't want surprising bugs and unnecessary code in my end result.
You can use an LLM, review the code and therefore avoid surprising bugs and unnecessary code in your end result.
Admittedly havent used deepseek v4, but v3 was so overhyped and bad that I'm reluctant to wasting my time on it.
Maybe you will inspire me to use it.
I did this and I use small local models as a productivity booster. It's been refreshing
Hints or tips on how to start with local models? I’m considering a new MacBook Pro and wondering if I should take that into account.
installing a local model gives you time to work on the important code and let the ai do the drudgery
So close to doing the same
That’s incredibly frustrating.
I’ve got a NixOS Qemu VM I use to run openclaw in. I had Claude help me set it up, and it runs local models on my own machine in a config based sandbox.
Why should Claude block or charge extra to work on that?
Why should Claude care if I have instructions for Hermes or OpenClaw in my project repos?
This fingerprinting is incredibly sloppy for how much access to a machine Claude code has.
Now you've learned the advantage of knowing how to do things yourself. When you depend on untrustworthy agents, you shackle yourself to their idiotic whims. Be careful who you partner with.
If it's just to set up a VM, how much would you even need to use? A couple of cents?
I run an OpenClaw VM and used Claude Code to build the VM scripts. The VM is connected to local llama.cpp, so OpenClaw and the models are running on my own physical hardware.
That is a huge red-flag. While I understand that they will do some policing/censoring, this is way beyond what I would consider acceptable.
They can have a different price plan for agentic stuff, but these things where they “accidentally” whoops match on specific keywords and trigger extra usage charges is giving a evil-microsoft-vibe
same vain as https://news.ycombinator.com/item?id=47952722 ?
@bcherny well need a bit more than a "Fixed" here... https://github.com/anthropics/claude-code/issues/53262#issue...Sounds exactly like what you’d get if you asked Vlaude how to detect OpenClaw usage.
They are trying to make a moat where no possibility of creating a moat exists.
It’s a huge mistake at the level of IBM trying to reestablish dominance over PCs by making MicroChannel the new standard; this failed horribly and cost IBM its market leadership and reputation.
MCA was technically better at the time, but the industry responded with EISA and VLBus which led to PCI and today’s PCIe.
Things like these (Google also banned me from Antigravity for briefly using an agent) and the massive quality swings made me cancel all 3 subs last week and resort to my local Qwen 3.6 only. Open models are already great and only getting better, and I really enjoy the privacy and consistency of a model I run myself.
I don't think anyone is questioning all the benefits of using local LLMs. Those are readily apparent.
I just don't believe for an instant that they're anywhere in the same ballpark of capabilities as running Opus or similar. My time is the most valuable resource. Opus would need to be SIGNIFICANTLY more costly and unstable for me to start entertaining local models for day-to-day development.
Perhaps whatever work you're doing makes this trade-off more sensible, but I struggle to see how that could be true. I'm averse to running Sonnet on a large amount of software engineering problems - let alone Qwen.
I think you'd be surprised, I find that the harness is what makes the real difference. I also prefer to be on the loop, actively guide and review. Local models are definitely much less autonomous as of today so if you need to be churning out code at speed they're probably not for you.
If you know what you're doing and prompt it correctly, local models are great. If you're just vibe coding and relying on the LLM to fill in all the gaps for you and basically build the software for you, yeah you need SOTA to deal with that.
But, you know,
Yet.
For now we infer through few weights, lossily; but then in full precision. Now I represent in part; but then shall I represent as fully as the data was sampled.
1 CorinthAIns 13:12
This is the future.
How much VRAM do you need to achieve decent performance?
I have a 64GB M1 Ultra dedicated to llama.cpp. I get 40 tok/s on a fresh session, decreasing slowly to about 25 tok/s at around 50% of the 256K context, then down to 20 tok/s or less beyond that, but I rarely let it go much higher and handoff instead. This is whith Qwen 36B A3B at 8Q without KV quantization. It's not super fast but perfectly usable for me.
Subscription models only work when marginal costs are low and/or there’s a good variety of usage that roughly averages out. Or, you need to be able to kick out abuse.
Unfortunately for those of us who just want to eat a nice filling meal at the fixed price all you can eat buffet of AI subscriptions, a minority of customers keeps paying for the all you can eat buffet and staying for hours and bringing containers to sneak food out when they leave. And they keep wearing disguises to try and evade detection.
It’s a losing battle for the provider, which ultimately means the subscription pricing model can’t work, which hurts the majority of customers that just want to use the system as intended and no longer have a subscription model available.
I have plenty of frustrations with Anthropic as a paying customer, but tbh is doesn’t strike me as all that awful, just some annoying collateral damage. I’d rather have that than no subscription model at all.
I wouldn't be surprised if the AI usage model moves towards a bidder/auction model. Set how much you'd willing to pay for your AI request, and they evaluate requests starting from the highest to lowest bids.
It definitely would make sense, especially if they are capacity constrained, but it’s also a losing PR move for whoever moves first in the space unless the big players all shift at the same time.
Is it just me or everybody finds the "charging extra" a bit vague? I don't deny it simply curious: how much?
Who remembers the Google of Eric Schmidt and "Don't Be Evil"?
The truth is that it doesn't matter what companies say, what they claim, what they do, and what their CEO says/claims/does.
It's just a matter of time until the shareholders will get the right CEO to maximize shareholder value.
People in the comments who want a statement or a "reorientation" or a commitment from Anthropic leadership are missing the principles of how capitalism functions. Shareholder value cannot be compromised. In every battle between morality and profit, values and profit, public good and profit, ultimately all things will mutate into a state to allow profit will prevail. Always.
There are no exceptions to this.
Several people at work, none use OpenClaw, had their limits jump immediately to 100%.
This is a reason to seriously consider changing providers.
I don't understand how, having access to Mythos and unlimited use, their solution to open harnesses is lazy string regex-style matching.
I saw a talk by Boris where he said, basically that Claude codes itself now. They have it automatically writing features and reviewing PRs, apparently. I suspect that much of the code has never been seen by human eyes within Anthropic.
lol so they aren't even good at using Claude
their CEO has been shouting from the rooftops that programming is dead. ofc that would ripple down the org chart and result in a culture of bad programming.
I wonder what happens if you ask Claude to solve the problem, and don't review it's answer properly..
they're just holding it wrong.. what model are they using? they should make sure they're on Opus 4.5+. That was a stepwise improvement and was when AI coding clearly became the futureₖₑₖ
https://xcancel.com/theo/status/2049645973350363168
why do people want to continue to use anthropic despite their shitty service? its not like they have some kind of lock-in as it is still new company and it has shown its color before we are stuck with it unlike google/meta etc.
Totally agree. This is why open source models and toolings are so important for the ecosystem. I would not want these companies decide what we can or cannot do.
When asking about Openclaw in normal Claude Webchat it very peculiarly denied knowing what that is.
Even when asked to search online it still gaslighted me about it.
When compute poverty hits these big labs it’s all going to be the same. The ping pong tables and drinks fridges disappear.
The only thing they can hope for is to maintain momentum and critical mass long enough to find ways to pay for all this or have Moores law make the average user request become economical.
possibly related, it errors if my working directory is a checkout of OpenCode. i was using CC to work on some patches for OC and had to work in a parent directory and then tell Claude to work on the files inside the "opencode" folder.
I tried to replicate this but Claude was already down https://status.claude.com/
It sounds like Anthropic is dangerously low on compute availability if they’re prioritizing these refusals as their OKRs.
I think it's obvious that they are critically lacking in compute capacity especially since OpenAI has committed billions to locking up all the future compute production.
And I don't necessarily think it's wrong for Anthropic to introduce QoS or throttling on users of their models. It's pretty much a necessity when offering public access to a scarce resource and it's been a common practice for decades.
What is the alternative? We just accept that it doesn't work half the time because the system is overloaded with molt bots?
I agree. If compute is the issue and pricing can't budge then something has to give.
They would have kept my business if they were honest and upfront. Instead they sold me something that worked well, broke it without warning, remained silent about it until enough people caught on, chose to do nothing, then proceeded to release a model that eats ~30% more tokens with no advantage over prior models.
If they chose to unbrick their model and offered what we had a couple months ago at a 50% hike, I would have been onboard. I've seen enough now of how this company treats its customers to continue using or recommending them.
Also, Codex works much better than CC now for anyone who happens to be on the fence.
The alternative is to price their product transparently. If there is too much demand and supply is limited: Charge more.
Anthropic wants to have their lunch (low apparent prices, increased market share) and eat it too (controlled costs, adequate production to serve the demand).
They're advertising themselves as a $5 All-You-Can-Eat buffet, but then aggressively and arbitrarily restricting admission, sneakily swapping out the high-quality ingredients for garbage-tier slop, and kicking out anyone who even utters the words "to go box" or "doggie bag".
Would you want to eat at that restaurant?
If they can't serve all their existing customers maybe they should stop accepting new customers until they can?
Maybe they could not sell more if they’re already exceeding capacity? What kind of apologism is this?
I cancelled my subscription so not really defending them myself but if all of their customers were humans who used it normally I bet they could serve everyone. It's when someone presses a few keys walks away and a bot uses tokens for 72 hours straight that it becomes a problem. Then people buy 3 accounts and do that for weeks at a time.
Could you do that as a human? Sure but you'd likely burn out after a couple of weeks. Also the human would probably use those tokens far more effectively and would not need as many. It's feels the same as someone installing a crypto miner on their servers in my mind. Abhorrent behavior.
I find it incredibly that after all the good faith Claude Code built during 2025 they are destroying users trust is such amateurish ways (same as hermes.md)
At least we can assume that Anthropic eats their own dog food. They use Claude to develop their software.
You say that like it's a gotcha. I think the fact that they reached 2B/mo in revenue by dogfooding cc is all the proof that one needs that this thing actually works. In fact it works so well that more people want it than they can serve. For months now they've been having issues when EU and US tz are both online at the same time.
> I think the fact that they reached 2B/mo in revenue by dogfooding cc is all the proof that one needs that this thing actually works.
That's a notable achievement, but let's have some balance... It's also responsible for the biggest self-own in software industry history by leaking their 1) crown jewels (i.e., source code) 2) the existence of their next model Mythos, and 3) their roadmap in a highly competitive market.
Eh... I personally think that having the keypads to enter a DC running on DNS served by that same DC is a bit more self-owning than leaking the source code of an app, but I get your point. It's obviously not perfect, but it's also obviously working.
Let's put this in perspective. Imagine it's 3 years ago, April 2023. Chatgpt has been launched for 4 months. We've all been using it, and writing poems in parrot talk or whatever. Someone tells you "In 2 years time there will be an app that lets you use LLMs to write code. It will be coded by humans for 3 weeks, then by humans + LLMs for 6 months, and then by LLMs mostly unsupervised. One year after that, they'll be making 2B/mo out of that app". Would you believe them? Not even the most maximalist, overhypers, AI singularity frenzied crazy people would have said that. And yet... it happened.
Is the reason they reached 2B/mo partially contributed by the fact that their users feel like they get unlimited use of it? If ‘feeling like it is unlimited use’ is a huge part that creates the 2B/mo, this change of limit might jeopardize it.
That being said, Anthropic can be diverting capacity to train the next model, and if it is significantly better, people would start flocking back again.
Everything works until it doesn’t.
The problem with slop is, nobody understands it. Nobody ever designed it, nobody really knows how it works. You’re just putting blind faith in the slop you’ve shipped.
It lets you be very quick, but if you’ve accidentally compromised all your data or bank accounts through the slop then you won’t know until you’re destroyed.
So what's next ... they are going to charge you a 30% commission on your sales for products build with their tools?
I find it interesting that I use Opus tokens and I have 0 issues.
Claude is bad for business....that is painfully obvious.
At this point I assume you are coping with having drank the koolaid and fired key staff believing claude will replace them...back when it was cheap....because nearsightedness affects decision makers much more during hype cycles......
i wouldn't be surprised if we see class action lawsuits from this given it's so easily reproducible by so many
Having had Claude Code jump to inserting juvenile and all-filtering regex to (attempt to) solve open-ended semantic natural language problems (-sigh- there's 12 hours of my life I'll never get back), I can absolutely imagine that this was someone trying to code up a "defense in depth" mechanic that was explosively insufficient after Claude Code (even Opus 4.6) made a series of faulty assumptions.
This one feels like prime space for Hanlon's razor: "Never attribute to malice that which is adequately explained by stupidity."
The hassle with the performance of these systems is that they're ~70% of the way to awesome. For advanced prototyping (my current job description), a fast 60% of awesome is groundbreaking and game-changing. For production and real businesses, that last 30% is a really, really important thing to figure out.
I hereby propose we rename the HN frontpage to "Claude Customer Support"
Highly relevant: https://en.wikipedia.org/wiki/Principal–agent_problem
(You're the principal, directing what to do, but your agent Anthropic has its own motivations that are not aligned with your will.)
Anthropic is losing a ton of goodwill by not being more honest about their constraints. They've been buckling under load for months, and instead of doing the most honest thing (keep weekly usage limits same, make 5 hour usage limits have surge pricing where the usage-cost of X tokens is scaled based on dynamic load), they're doing a lot of hacky things to try to get a similar effect. I suspect they feel the optics of being honest would be too bad, so instead it's a slow bleed where they piss off users one by one
do they literally just have a regex match for all of their competitor harnesses?
nah, it's probably worse: it could be some system prompt for their models...
But Peter Steinberger said that openclaw was “fully supported” with a subscription through claude -p.
Do these refusals still happen if you’re using an API key instead?
So I suppose Anthropic lied to him?
In response to this he said "WAT"
Ok I am usually defending Anthropic, but it seems like this OpenClaw and Hermes ban was implemented incredibly poorly; it looks like a simple regex.
Didn’t they think about “we need to make sure Claude Code is never banned” ? Could have been as easy as including some Claude Code specific prompting traits (tools, system prompt, whatever) in there and automatically whitelisting it.
Is it foolproof? No. Will it avoid banning legit users? Absolutely.
First do the first large sweep, then see what still falls through, then ban those.
It really seems they were panicking due to capacity and there was very little oversight with all this.
I’m not affected but pretty disappointed.
Why would you defend Anthropic at this point after all their antics and their behaviour over the past 6 months?
They do not care about us.
I love their vibe coded "anti-abuse" systems :D
If they're gonna vibe-code all these arbitrary rules, they should at least release the source code so we can figure out how to work around them!
Oh come on Anthropic, just admit straight away that any other pricing than usage-based is completely unsustainable and is being phased out.. maybe doing it once but officially could save you some brand damage.
what a company with really bad customer practices. I'm really glad I moved entirely to open source models. if you're disgusted by these practices as I am, I really recommend you use opencode (or any of the other 20 agents) and the GLM 5.1, or Kimi K2.6 or Deepseek V4 Pro models. You will be shocked how effective they are.
haven't used claude in about 2 weeks and I do not miss it.
Interesting people talking about whether they should be "defended," here or whatnot, and all of that strikes me as wildly naive.
They have a business model that's more or less known, and that includes THEIR AI model(s) that they get to put out there however they want. I don't like it much at all, I actually sort of like the idea that they "owe" more because they probably "stole" a bunch of stuff to get the thing going.
But I mean, don't be mad, be proactive. Anthropic is going to try to Microsoft this in whatever way possible, and we all see that the numbers don't really add up.
Asking them pretty please to be nicer, meh. Let's figure out better, and more free-software-like ways to do this.
the most relevant person on this industry Theo - t3.gg /s
:3
I think that’s an ok move, definitely better than canceling code on pro users for example, I would support to even have a new pricing tier only for openclaw, so they don’t ruin the usage on others. I noticed the ones who use claude code usually are software developers or sysadmins, meanwhile most openclaw ones are your average HR stacy and lazy middle managers, so yeah, it should be a separate tier for them.
I think the pricing tier for open claw should probably just be the per token API one?
openClaw does so muhc more then Claude code tbh, running 9 agents from the one machine, schedual some tasks, add some personal personas for each agent, claudeCode (which i like alot) is on rails, openClaw is full openworld.
rate the analogy plz..