I've run into this, and I highly doubt I am one of the more extraordinary users. I have delays between working with it, don't have many running at once, am running on smaller codebases, etc. Yet just a few minutes ago I hit a quota. In the past I did far more work with it without running into the quota.

I emailed their support a few days ago with details, concerns, a link to the twitter thread from one of their employees, and a concrete support request, which had an AI agent ('Fin') tell me:

> While our Support team is unable to manually reset or work around usage limits, you can learn about best practices here. If you’ve hit a message limit, you’ll need to wait until the reset time, or you can consider purchasing an upgraded plan (if applicable).

I replied saying that was not an appropriate answer.

You're absolutely right re the lack of transparency and accountability. On one hand, Anthropic generates good will by appearing to have a more ethical stance then OpenAI, and a better product. On the other hand, they kill it fast through extremely poor treatment of their customers.

If they have a bug, they need to resolve it: and in the meantime refund quotas. 'Unable to' - that's shocking. This is simple and reasonable. It's basic customer service. I don't know if they realise the damage their attitude is doing.

It is also interesting to observe that your most valuable accounts in this kind of pricing model are the ones that are least used and therefore are not confronted by the limits. Heavy users canceling their accounts in frustration is a win for Anthropic not a punishment, at least a short term.

Once you get used to using claude as an abstraction layer you start getting pretty reckless with it.

My organization has the concept of "premium models" where our limits reset every month. I hit my limit pretty quickly last month because I was burning tokens doing things that would have been a simple bash loop in the past - all because I was used to interfacing with Claude at the chat layer for all my automation needs and not thinking any more about it.

This feels a lot like the same playbook we’re seeing with dynamic pricing in retail, just applied to compute instead of products. You never really know what you’re getting, and the rules shift under you.

What makes it worse is the lack of transparency. If there were clear, hard limits, people could plan around it. Instead it’s this moving target that makes it impossible to trust for real work.

At some point it stops feeling like a bug and starts feeling like a pricing experiment on users.

They keep running experiments like free $50 in extra use credits or 2x usage outside certain windows where inference is very slow. You can’t help but think this is all a slowly boiling the frog experiment. Experimenting how much they can charge.

Are they going to pay back if subscription was payed but token limit was less than advertised? Is there some tiny text somewhere preventing just suing or pulling money back with credit cards?

I suspect that Claude had a bug that undercounted tokens and they fixed it.

Working as intended? They openly state that how quickly your limit is reached depends on many factors (that you don't know) as well as current load on their systems.

Could just be that usage has gone up.

Everyone on my team has been running into this, including the super users on the Max plan and the skeptics who only use it every few days. The quota is going way faster than it did before, sometimes a single prompt will eat up a third or more of the session quota.

It's more structural than a simple bug. The cch mechanism embeds a sentinel string (cch=00000) in the serialized request body, which gets replaced by a hash computed in Bun's native runtime (compiled Zig). The hash covers the entire JSON body, so it functions as request integrity verification. The fragility comes from doing string replacement on the payload itself. If cch=00000 appears anywhere in the conversation history — discussing billing code, quoting source, even this thread — the replacement can hit the wrong occurrence, corrupting the body and invalidating the prompt cache. This is a known anti-pattern in protocol design: mixing your signaling channel with your data channel. The standard approach is to put integrity hashes in a request header (e.g., HMAC in Authorization), keeping the signature completely outside the body it covers. That way no payload content can collide with the signing mechanism. The choice to embed it in the system prompt was presumably to avoid API surface changes, but it created a category of bugs that can't be fully fixed without moving the hash out of the body.

I'm not sure this is the issue. I asked Claude Code a simple question yesterday. No sub agents. No web fetches. Relatively small context. Outside of peak hours. Burned 8% of my Max 5x 5hr usage limit. I've never seen anything like this before, even when the cache is cold.

> BUG 2: every time you use --resume, your entire conversation cache rebuilds from scratch. one resume on a large conversation costs $0.15 that should cost near zero.

I use it with an api key, so I can use /cost. When I did a resume, it showed the cost from what I thought was first go. I don't think it's clear what the difference is between api key and subscription, but am I believe that simply resuming cost me $5? The UI really make it look like that was the original $5.

Nothing devious, but is Anthropic crediting users? In a sense, this is _like_ stealing from your customer, if they paid for something they never got.

your linked bug is a cherry pick of the worst case scenario for the first request after a resume.

While it should be fixed, this isn't the same usage issue everyone is complaining about.

That bug would only affect a conversation where that magic string is mentioned, which shouldn't be common.

Anecdotally when Claude was error 500'ing a few days ago, its retries would never succeed, but cancelling and retrying manually worked most of the time.

That is a summary and a picture of https://old.reddit.com/r/ClaudeAI/comments/1s7mkn3/psa_claud... it looks like?

After they force OpenCode to remove their Claude integration, and the insane token hogging, I also cancelled my subscription.

I have had the exact same experience (like super uncanny with prices etc). And now feel like I can only use my Claude subscription for the most basic issues. I’m getting range anxiety.

This reminds me of the early days of cell phones. Limits everywhere and you paid for it by the kilobyte. Think at one point I was paying 45c per text message. I hope this gets better and we do not need gigawatt datacenters to do this stuff.

Yeah, I've been juggling some patches to opencode to help me see where my codex usage limits are at. As of a month ago, that information was not visible on the ChatGPT web UI.

You just work till suddenly the AI dumps you out, and sit there wondering how many hours or days you have to wait. It's incredible that this experience is at all ok, is accepted

Tool calls (particularly fetching for context) eats the context window heavily. I explicitly send MCP calls to sub agents because they are so “wordy”.

I haven't found Kimi to be all that good, but GLM 5.1 I find to be better than Opus 4.6 most of the time in web dev. Opus' only advantage is it's a bit faster. If you can't access GLM 5.1 (not fully released yet) try 5.0. It was better than Opus sometimes too.

I have a GLM Code subscription and it lasts much longer than Claude Code.

I use Pi agent so I use all agents in the same harness.

That is wild to me. On Claude Pro, I have never hit the limit, and I use Claude every day.

I’ve dumped claude few months ago for gemini. Maybe my problems are too trivial, but it’s same if not better with added benefit of being much faster. I’d say 95% of my work (20-30hrs per week) is done by it and I spend less than $50 per month.

As long as they keep losing money and are reliant on investments to pay their operating expenses, they are going to be thrashing about in search of a sustainable business model and I don't blame them.

Oh they’re very competent. They’re running a scam.

Theyrr a business abd tgeyre burning cash. Their goal is to find profit and that typical grades towards whales.

We have decades that say this. Few businesses survive by low margins on wide audiences.

This is of course subjective, but I would give a lot to have an alternative to Claude Code and the Claude models, but there just isn't anything comparable that works well in an integrated manner for agentic coding.

It's not like I haven't tried. Gemini CLI is still trash (it's probably a bit better now, but I still can't see the edits it proposes, well, etc.). I tried OpenCode, the whole experience was frustrating: the models give up mid-task, they run rampant with actions, the CLI does not offer the level of control and customization Claude Code offers, etc.

I've also tried the other major tools: Codex, Cursor, Cline, Aider, and others, nothing works for me. You are surprised people stick to Claude Code, I am surprised people bother with the other tools.

Maybe it has something to do with how I use the agentic tools: I use the CLI almost exclusively, rarely using the IDE (unless I want to actually code myself). I also almost always approve each and every edit. As such, my number one concern is for the tool to provide me with proper control in a simple and reliable manner: I want a rich permission system that works, and I want to see each proposed edit very clearly in an ergonomic diff format. I want to be able to type, recall, and edit my commands easily too. These are things Claude Code excels at that the other just don't.

The best I've been able to do is to use third-party routers to enable me use Claude Code with almost-SOTA models, and this is the approach that shows the most promise. I'd hate to be beholden to Anthropic's shenanigans.

Also the biggest models are not always the best depending on what you're doing. 4.6 Sonnet is a decent model that can handle most coding tasks, and even 4.5 Haiku is fine for simple, well-defined tasks.

Using 4.6 Opus for simple things is not only wasting tokens, it's also slower. Sonnet will get a lot of tasks done in half the time for less than half the money

This has largely been my experience. Can’t tell the difference between Claude and kimi

[dead]

You have better luck just cancelling your subscription. Claude is becoming too expensive for what I use it for. I don't want a refund. Maybe I just realized that I do not need a coding agent.

I definitely learned to plan out my projects more using LLM's, but in that case im 80% there. I might hit a roadblock or two, but if that means I don't have to guide an LLM then I'd prefer that.

How do you even make a stink? I haven't found an easy way to find a human.

When I got my Google AI Ultra, I could run it morning to evening at opus 4.6. One month later I started hitting 5h limits when I was nearing my 5h window. 2 weeks after I hit my 5h limit 30 min into the morning. Cancelled my sub even quicker.

Free is free. Want more, fork over money.

If this is reasonable or not is pretty hard to judge without any info on that "ONE" task.

Even worse, the open weight models are practically indistinguishable from the closed ones. I just don’t see why you’d pay full price to run Claude when you can pay 10x less to run Kimi. There are already loads of inference providers and client layers.

Without heavy collusion or outright legislative fiat (banning open models) I don’t see how Anthropic/OpenAI justify their (alleged) market caps

> the cost $1000-2000/month or more for coding. Every programmer will be forced to pay for it because it's only a fraction of their salary (in the US anyway) and it's the only way the programmer will be competitive.

I routinely match or beat Claude with regards to speed, I often race it to the solution because Claude just takes so long to produce a usable result.

Staying competitive doesn't mean only paying an AI for slop that often takes longer to produce. AI is a convenience, it is not the only way to produce code or even the most cost effective or fastest way. AI code also comes with more risk, and more cognitive load if you actually read and understand everything it wrote. And if you don't then you're a bit foolish to trust it blindly. Many developers are waking up to the reality of using AI, and it's not really living up to the hype.

I get the same. Work has shifted to being agentic first - and whenever I use anything other than Claude Opus it seems that the model easily gets lost spinning its wheels on even the simplest query - especially with some of our more complex codebases, whereas Opus manages to not only reason adequately about the codebase, but also can produce decent quality code/tests in fairly short order.

Oddly though, when using at home I'm using Sonnet via the standard chat interface and that, whilst it will produce substandard code in its output is still reasonably capable - even in more niche tasks. Granted though that my personal projects are far simpler than the codebase I handle at work.

An aside: https://www.buchodi.com/chatgpt-wont-let-you-type-until-clou...

It seems that anthropic has added something similar to their browser UI because just in the last few days chat has become almost unusable in firefox. %@$#%

Its a question of price, quality and other factors.

If my company pays for it, i do not care.

If i have a hobby project were it is about converting an idea in my spare time in what i want, i'm happily paying 20$. I just did something like this on the weekend over a few hours. I really enjoy having small tools based on single html page with javascript and json as a data store (i ask it to also add an import/export feature so i can literaly edit it in the app and then save it and commit it).

For the main agent i'm waiting for like the one which will read my emails and will have access tos ystems? I would love a local setup but just buying some hardware today costs still a grant and a lot of energy. Its still sign cheaper to just use a subscription.

Not sure what you mean though regarding speed, they are super fast. I do not have a setup at home which can run 200-300 tps.

You are not crazy, you are just waking up from the SaaS delusion. We somehow allowed the industry to convince us that paying $20/month to rent volatile compute, have our proprietary workflows surveilled, and get throttled mid-thought is an 'upgrade'. The pendulum is swinging violently back to local-native tools. Deterministic, privately owned, unmetered—buying your execution layer instead of renting it is the only way to build actual leverage.

I'm burning through pretty fast with context sizes of only 32-64kb. I regularly clear when I change topics.

A simple "how do I do x" question used 2% of my budget.

I paid extra and chewed through $5 in a few minutes of analyzing segments of log files.

At this rate it's not worth the trouble of carefully managing usage to avoid ambiguous limits that disrupt my work.

If that's the way it is in order for them to make money, that's fine - but I need a usable tool that I don't have to micromanage. This product is not worth it ($, time) to me at this rate.

I hope it changes because when it works it's a great addition to my tools.

> on a single entity

Contrary to the popular opinion here, there are other services beyond Claude Code. These usage limits might even prompt (har har) people to notice that Gemini is cheaper and often better.

For a second I hoped you were gonna comment on how LLMs are going to rot out our skillset and our brains. Like some people already complaining they "have to think" when ChatGPT or Claude or Grok is down.

Oh well.

There's so many different models, from hosted to local and there's almost no switching cost as most of them are even api compatible or supported by one of the gateways (Bifrost, LiteLLM,...).

There's many things to worry about but which LLM provider you choose doesn't really lock you in right now.

So, like, GitHub then?

I don't get this pov, maybe b/c I'm not a heavy Claude Code user, just a dabbler. Any LLM tool that can selectively use part of a code base as part of the input prompt will be useful as an augmentation tool.

Note the word "any." Like cloud services there will be unique aspects of a tool, but just like cloud svc there is a shared basic value proposition allows for migration from one to another and competition among them. If Gemini or OpenAI or Ollama running locally becomes a better choice, I'll switch without a care.

Subscription sprawl is likely the more pressing issue (just remembered I should stop my GH CoPilot subscription since switching to Claude).

It should be abundantly clear that depending on a single entity will screw you royally, but obviously we don't learn from the mistakes of others. We are condemned to repeat history because we don't know it.

How can automatic slop-prevention be a disaster? It's a feature.

if you rely on the black box of bullshit... you deserve your own fate.

If you want to still use APIs, I like OpenRouter because I can use the same credits across various models, so I'm not stuck with a single family of models. (Actually, you can even use the proprietary models on OpenRouter, but they're eye-wateringly expensive.)

Otherwise you should look into running e.g. Qwen3.5-35B-A3B or Qwen3.5-27B on your own computer. They're not Opus-level but from what I've heard they're capable for smaller tasks. llama.cpp works well for inference; it works well on both CPU and GPUs and even split across both if you want.

i would recommend getting an API account on fireworks, this is ZDR and typically the fastest provider.

otherwise check the list of providers on openrouter and you can see the pricing, quantisation, sign up directly rather than via a router. ensure to get caching prices, do not get input/output API prices.

GLM 5 is a frontier model, Kimi 2.5 is similar with vision support, Minimax M2.7 is a very capable model focused on tool calling.

If you need server side web search, you could use the Z AI API directly, again ZDR; or Friendli AI; or just install a search mcp.

For the harness opencode is the normal one, it has subagents and parallel tool calling; or just use claude code by pointing it at the anthropic APIs of various providers like fireworks.

Just install ollama.

And no, they're not as capable as SOTA models. Not by far.

However they can help reduce your token expenditure a lot by routing them the low-hanging fruit. Summaries, translations, stuff like that.

We offer multiple SOA models at https://portal.neuralwatt.com at very generous pricing since we have options to bill per kWh instead of per token. Recipes for your favorite tools here: https://github.com/neuralwatt/neuralwatt-tools