[BUG] Weekly Usage Limits Making Claude Subscriptions Unusable - This IS a bug.
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
Reopening the following issue as it was closed erroneously.
Summary
Users across Pro, Max x5, and Max x20 subscription tiers are reporting that weekly usage limits are being depleted in 1-2 days of normal use, making paid subscriptions effectively unusable for 5-6 days per week.
Impact
Pro users: Hitting weekly limits in 1-2 days
Max x5 users: No meaningful improvement over Pro
Max x20 users: Still hitting limits quickly despite $200/month subscription
Users report cancelling subscriptions and switching to competitors
Detailed Issues
- Weekly Limits Reached in 1-2 Days
Multiple users report exhausting their entire weekly quota within the first 1-2 days:
"i used it a little firday a little on saturday and maybe 10 minutes this morning and i hit the weekly limit already"
"I burn through the whole damn quota in like ONE OR TWO DAYS"
"hitting weekly limits as a pro user in a day and a half"
"The weekly limit is now reached in one day"
- Higher Tiers Not Providing Expected Value
Users who upgraded to higher tiers report no meaningful improvement:
Max x5: "5% used from ONE message on a 5X Max account"
Max x20: "I upgraded from Max x5 just yesterday, but the weekly limit is still a problem. The rate at which my Opus 4.1 usage limit depletes is no different from Max x5"
Max x20: "$100 Max is the new PRO- PRO was just a trial"
- "Bait and Switch" Concerns
Users feel misled by subscription changes:
"This feels like I've been bait and switched TWICE now"
"Whether you call it bait and switch, predatory pricing, or consumer deception, it's just plain unfair"
"I'm literally paying for a service I can't use"
- Opacity in Limit Resets
Reset times reportedly change: "they postpone the reset time silently after a week. My weekly budget should be reset at Wed 10pm. Yet after resetting, the next reset time becomes Thu 11:59am"
- Professional Use Cases Blocked
Coding, creative writing, and data extraction workflows are being interrupted mid-project
"5 hour limit is okay, you can plan around it... Cutting off usage mid work-week is like losing your top developer"
"6 days of production a month is not worth your subscription"
User Sentiment
Multiple users reporting subscription cancellations
Switching to competitors (ChatGPT, Gemini, open source models)
Comparisons to ChatGPT Plus: "On chatgpt plus i NEVER had to worry about hitting a limit"
Use Cases Affected
Software development/coding
Creative writing
PDF data extraction
General professional tasks
Research and analysis
Requested Changes
Remove or significantly increase weekly limits - Daily limits may be acceptable, but weekly limits compound the problem
Make higher tiers meaningful - Max x20 at $200/month should provide substantially more capacity
Transparency - Clear, consistent information about limits and reset times
Usage meter accuracy - Better visibility into consumption rates
Business Impact
Users explicitly stating:
"Canceled my subscription yesterday"
"I just cancelled my plan, looking for an altnertaive"
"I'll be going back to the API after this subscription ends"
"Just downgraded from Max X20 to Max X5, not worth it paying more"
Date Range
Complaints escalating since Claude 4.5 release (early October 2025)
Priority
CRITICAL - Directly impacting subscription retention and revenue
What Should Happen?
Fix it. Change or correct weekly usage limits.
Error Messages/Logs
Steps to Reproduce
Use claude on any paid plan.
Claude Model
Not sure / Multiple models
Is this a regression?
Yes, this worked in a previous version
Last Working Version
_No response_
Claude Code Version
1.0.93 (Claude Code)
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
PowerShell
Additional Information
Worked fine and was usable until someone messed up their job and failed their customers.
21 Comments
I have seen that my sessions on claude from powershell are being faster consume than before. what I was doind before in 3-4 hours, now It last maybe 20-30 minutes before reaching tokens and autocompact do the magic again....
I'm a 5x Pro user, I dont reach my limit per week easily, and havent reach yet, but need to wait for claude to compact each session every 15-20 minutes, even 10 minutes after just doing few commands, is annoying...
This need a fix.
I also noticed a big regression in usage with a regular Pro plan. The weekly limit did reset on Monday at 11 AM, it's Wednesday at 00:00 AM, and I already got a message that I am approaching the weekly limit. This is kind of unexpected and frustrating.
From favorite to unusable in just two weeks
For over 14 months Claude Pro was my daily driver and easily the best LLM I’d used. I happily paid every month and built extensive workflows in Projects.
That changed dramatically in the space of the last two weeks (November 2025).
Despite a freshly reset weekly limit:
Worst of all: if I cancel Pro, two years of carefully organized projects and conversations instantly become inaccessible. My own data is locked behind an active subscription.
The model is still excellent, but the implementation of the limits and the way they’re enforced have made Claude Pro unusable for any serious or sustained work.
yes, i thought i was just me but all week my context and usage gets used up in very quickly. after a few prompts - very little actual coding or reading of code or documents, my context window is used up and it compacts yet again. Its beyond ridiculous and very unusable.
if it doesn't get fixed i will have to move on to another LLM
UPDATE ---
I removed my comment above.
I noticed something different, there was an update how Claude shows limits, this morning, it was showing like this:
<img width="1079" height="304" alt="Image" src="https://github.com/user-attachments/assets/5ccc1659-f79f-460b-a130-f37439099323" />
And now is showing like this
<img width="466" height="242" alt="Image" src="https://github.com/user-attachments/assets/b4e2b359-50d1-4c9a-ae94-ebe3609f30fc" />
Note that, previously was used and now is left.
I have opened the claude code few days ago and haven't closed until now, after I opened in another terminal and closed and opened the one I had opened for few days. It started showing the difference, (even without updating it).
Make sure this is not what is happening to you.
You can see the same on UI
<img width="918" height="464" alt="Image" src="https://github.com/user-attachments/assets/175acf23-99b6-450a-bc48-e5f91b6e1da8" />
Guys,
do you have global Memory setting enabled in Claude>Settings>Capabilities? If so, try turning it off.
Same issue on Claude max + Claude Code v2.0.65. Quota depleting 2-5% per simple question since Dec 10 crash. Was working fine for 2 months on same 145k line codebase. Only since yesterday 11am, I didn't change anything from the config.
MAX 5x customer ($100/month) reporting identical issue since Dec 10, 2024.
Facts:
Support:
AI bot closed my tickets twice. Finally escalated after I explicitly mentioned "billing dispute" and "refund". Still waiting for human response after 10+ minutes.
Status:
Service completely unusable at $100/month. Evaluating alternatives.
Yesterday I was working as usual. Not doing a lot of work. Claude didn’t even compact as expected.
Example:
Tuesday 9, Claude compacted over 30 times in the session and everything was working “fine”
Wednesday 10, Claude compacted few times and alert reaching quota appeared. I continued working as normal In one session I reached the quota. I mean. Claude said 10% before autocompact. Alerts says my quota was at 90%.
Then Claude said 7% before autocompact. Alert said my quota was 97%.
Then Claude said 5% before autocompact. Alert said I completed my quota. To upgrade to Max plan, buy tokens or wait few hours.
I learnt what was social life again leaving early office. But the way quota system was working yesterday is not normal.
I decided to track my Claude Code usage, which I rely on most. I measured my productive time from zero usage until I hit the weekly limit, and the result was shocking: only about 24 hours of continuous use. Considering a month has four weeks, and I hit the limit after roughly 24 hours, my $20 monthly Pro subscription effectively provides only four days (96-hours) of usable time. A full month's subscription for a mere four days of access is, frankly, outrageous and completely unacceptable.
This situation is unacceptable. The five-hour limit was manageable. The weekly limit is an unnecessary and unwelcome addition.
I also believe that the 5-hour limit was manageable and, at some point, not the worst idea, but the weekly limit is really something Anthropic implemented that is actually broken or intended to limit users as much as possible.
To be honest, I had no issue with the 5-hour limit because I saw it as a "break." However, the weekly cap is literally a scam protected by "terms, conditions, and usage policy." It's humorous.
I've basically almost given up on Claude within a week of purchasing a Pro plan.
Claude pro plan in VScode sidebar is essentially unusable right now, if I give it more than 1 or 2 reference documents, it chunks it's whole 5 hour usage in ~10 minutes.
A task that takes a few percent of a ChatGPT (Codex, VScode sidebar) usage, can just wipe my whole 5 hour usage without even being close to finishing the task, which then leaves me in limbo for 5 hours, sometimes more than 10 hours, because Claude can't finish it in '2 x 5 hour window limits'.
1 simple question ("why are Claude 5 hour limits so low compared to Gemini or ChatGPT pro accounts?" - this question asked in the browser via claude.ai) can use up 10% to 15% of the 5 hour window.
I've tried being 'selective' with what I ask it to do, but I'm just not sure it can finish anything it starts, and will leave me having to wait another 5, maybe 10 hours, which puts me off using it TBH.
Yea. Too bad they would rather rip their customers off and steal from them. Sad to see. They were great, and now they are getting to be useless, which is horrible. Anthropic hates money, as so many many people have left over their greed. Nothing has changed for me and I am still having issues.
I suggest we try to tank Anthropic's sock price. That will receive a reaction.
No one will care. As long as elite developers from around the world rely on Claude Max for over a year and continue to do so. There are entire startups built and based on Claude.
@radeksalomon Does that really make such a huge difference? If it does, that feature is completely useless and nobody can ever use it for anything meaningful. Why does it even exist if you cannot use it for in real world scenarios?
I purchased a Pro plan only to test if Claude Code is a real alternative to OpenAI Codex. I don't like OpenAI (or Sam Altman), I would love to switch to another company. So I really would like to love Claude... but they are making it really hard to do so (website is buggy, VScode extension is ugly and buggy, real world usage limits are a ridiculously low)
I did a bigger coding tasks yesterday and within less than hour, over 30% of my Session limit was gone. Two more of those tasks and I hit my 5 hour limit and can wait maybe 3 hours before I can continue. And I only used Sonnet, I didn't even use Opus.
So for comparison, I made a copy of the project, reverted all Claude changes and gave Codex exactly the same prompts (copy&paste) to perform exactly the same tasks and Codex did perform equally well on all tasks but that ate about 7-8% of my session limit.
And note that the session limit you get with Codex is way smaller (way less tokens) than with Claude Code. Also note that in Codex I used the latest (and most expensive) model. Sure, my Codex subscription is more expensive (23 € for Codex vs 18 € Claude Code, 15 € if you pay a year in advance, all without tax), but I can do so much more with it before running into any limits. And it's about 2-3 times faster than Claude Code, so I really can do much more with it in the same amount of time.
I really hope that there is a bug because if that is by intention and disabling Global Memory does not fix it (which I doubt), Claude Code is unfortunately not a real alternative to Codex. And that is pretty sad as it means I'm stuck with a company I truly dislike.
BTW, this should not be tagged
platform:windows, since I can see the same issue on macOS.Hey, I think there are a few things worth considering before writing off Claude Code:
The comparison isn't quite apples-to-apples:
You used Sonnet (the "everyday tasks" model) on a Pro subscription, then compared it against Codex's most expensive model. If you look at Claude's model selector, it literally says:
Complex coding tasks = ambitious work. Sonnet can code, but it's not the recommended model for heavy development sessions.
Pro vs Max matters:
Pro has stricter rate limits AND no access during high-demand hours. I've been using Claude Code for about 6 months now, and back when I was on Pro (around November), I had to be extremely detailed with my prompts or I'd burn through tokens fast. After upgrading to Max, the experience improved significantly.
On model versions:
If you were using 4.6, that might also be a factor. 4.6 tends to be more "creative" (read: it can hallucinate or go on tangents more), which consumes extra context. Many of us actually prefer Opus 4.5 for coding because it's more grounded and less prone to overcomplicating solutions.
That said:
I won't pretend the limits aren't a known issue — they are. But comparing Sonnet on Pro against the top-tier Codex model on a pricier subscription isn't a fair benchmark. Try Opus 4.5 on Max before making a final call.
Just my 2 cents from someone who's been through the learning curve 🤷♂️
I didn't use Opus, because I read everywhere that Opus will use up your tokens even faster than Sonnet and Sonnet is good enough for everyday tasks and it is also the default model of the VScode extension.
If Sonnet is not suitable for programming, then why is it the default model in Claude Code? It's not like I actively selected it, I just didn't change it.
And my complain was not about the coding quality, not at all. Sonnet wrote excellent code for an AI (nothing crazy, following code style, doing exactly what I asked for), my complain was about eating up too much of my limits for that, so how would it have helped to use am model that will eat up the limits even faster according to other users and even according to Anthropics' own documentation?
I also don't use High or Very High reasoning with Codex for the very same reasons that it is very slow and eats up a lot of my limit, I usually only use Medium reasoning, which may produce worse results (or definitely will) but it's usually "just good enough for everyday use".
Again, I don't understand what Benchmark you are talking about. The benchmark I performed was token usage. Not speed, not capabilities, not code quality. Codex has a smaller token limit and I used a more expensive model. Claude has a much higher token limit and I used a less expensive model, so yes, the benchmark was unfair but only towards Codex, as Codex should have lost it big time; but it didn't.
And if Pro has no access during peak hours, than Claude just got the way worse deal, because that's an issue I've never seen with Codex at all since the first day I used it. Also how would comparing Max for 90 € be a fair comparison against Codex for 23 €? Basically all you did with your post is hardening the fact that Codex gives you more value for you money (more coding tasks per €) and is the way better overall deal.
Same pattern here — ran my last 7 days through a small analyzer I wrote and got $5,163 burned on Max $200, 100% Opus, 11.2K calls. The weekly cap on Max $200 is community-estimated around $320 of token value, so I was running ~15× my subscription value before getting cut off. Subscribed for "unlimited-ish" and getting invoiced-ish.
The part I didn't expect: of that $5,163, the analyzer flagged ~$2,062/week as Opus calls that would've cleanly run on Sonnet (simple edits, refactors, boilerplate). So while this thread is right that Anthropic's silent changes are the headline bug, there's a secondary lesson — the tooling ecosystem doesn't make it easy to see your model-usage split before you hit the wall.
For anyone else who wants to pull their numbers for this issue: https://genie-j.github.io/burncheck/ (single HTML, client-side, nothing uploaded). Not promoting it — just there's a gap where people in this thread are frustrated but can't produce concrete data to push back with. Takes 30s to generate a shareable card of your own burn.
In practice, we don't need hard data to fight back with. Since 2024, I've been using Anthropic models on the same Pro plan. It's easy to see that your productivity was much higher back then than it is now.
Are people leaving OpenAI for Anthropic? I did the opposite. The price is the same for performance that is very close and well-balanced. The only downside is that it is slower, which is an acceptable trade-off.
Add me to the angry list, a single prompt and it eats my limit. I switched off everything I can that people say turn off, i start brand new chats, I moved down from opus to sonnet, I work in the middle of the night in USA off peak hours, I tried it ALL, and a single prompt to edit a single word ona wordpress site ate my entire session limit, broke mid file upload left my production site down with half a template file truncated in it so now its not just go fix a dangling semicolon, half an entire critical file is missing on my server and its five hours till I can even ask claude to let ME replace the file, because I do not have the file, because its in Claude's session limited half finished response somewhere.
Angry? To say the least, WE PAY for this? And to have to go back to Atman's pile of steaming crap because it at least delivers what you paid for? That stings worse, cause ALTMAN sucks as a human and his company sucks as a company for reasons that aren't even about how the LLM actually works (as bad as GPT is... its not the reason to hate altman and openAI, there are much better reasons) .