Obscenely Long Rate Limits (Over 24 Hours) and Possible Distillation False Positive

[pepperymilkcap]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

Copilot in GitHub

Body

When did these become a thing? I'm a Pro+ customer who consistently pays for hundreds worth of extra Claude Opus 4.6 credits per month. I might consider cancelling if this isn't a bug because I cannot afford to be hung out to dry for days at a time.

I tried switching to Opus 4.5, then Sonnet, but nothing budged.

After a while I attempted to start a chat using the web interface, only to be met with an suspicious half-accusation of scraping:

I've opened a support ticket but was wondering if anybody else has had similar issues? I frequently use Copilot in 1-3 windows simultaneously and used it under a VPN for a quarter of a hour or so earlier.

[LegionGPOS]

66 hours!!!!! paying enterprise prices for a single person on copilot pro+ you all are F*ing insane!!

[LegionGPOS]

make syou wonder what teh actaul rate you get is...for the week....

[pepperymilkcap]

Have you opened a ticket with support? I've just opened one. Hopefully this'll be fixed and I can get some work done tomorrow.

[LegionGPOS]

yeah i did.

[LegionGPOS]

auto mode seems to work but who knows how long that will last and again when you're paying, being forced to use it is crazy.

[suacasarendemais-png]

Estou com o mesmo problema. caso não haja solução, que ferramenta pode substituir o vscode e copilot? como ter uma segunda alternativa?

[sgaabdu4]

I'm having the same issue ^

[mayank4574]

Same issue happened to me. It’s due to rate limiting.
Fix close extra tabs, disable VPN, logout/login, and wait for cooldown.
After that, it worked fine. Not a bug — just strict limits.

[LegionGPOS]

"Not a bug" no, deliberate action they admitted to and a "contributor" just reiterated.

[pepperymilkcap]

No solution as of yet. Resorted to auto mode and am getting sub-par code quality.

[LegionGPOS]

yeah and even worse than that i had one system that couldn't use anything but gpt, so i'm using antigravity to sort it out and see if i can get it to use auto afer fixing the extension; i digress, i guess i'll be finding a suitable intermediary and limiting my github usage at this point. they definitely got greedy overnight with that and i know folks are making a killing off the api's for this. i have 22llm's i'm bound to break one free for my own copilot.

[LegionGPOS]

no way i'm paying over $100 to get this kind of nonsense.

[LegionGPOS]

Insult to injury...why is teh cli even longer? yeah i definitely need to make better moves now. now auto for the cli either.

[Zero3K20]

Auto mode doesn't work for me.

[pepperymilkcap]

Perhaps related?
https://github.blog/changelog/2026-04-10-enforcing-new-limits-and-retiring-opus-4-6-fast-from-copilot-pro/
I don't think I've seen any other service that punishes paying customers for spending more.

[LegionGPOS]

anthropic was already taxing them from what i read so this makes sense, but as you said, we get punished for cashing out...which is crazy. What's even worse is the "upgrade your plan". Uh, you only have pro+ for regular users...we can't go any higher and this crap here is not how you get folks to look into your plans when you have so many separate payment situations for everything already. and again, there's the overbudget feature that is now completely useless with this change.

[Incognitiv]

Sorry, you have been rate-limited. Please wait 18 hours 47 minutes before trying again or consider switching to Auto. Learn More

Server Error: Sorry, you've exceeded your weekly rate limit. Please review our Terms of Service. Error Code: user_weekly_rate_limited

Of course I have, that's why I'm paying you (GitHub) to get rid of the limits. I won't pay more to increase my "weekly rate limits" as well if I'm just a hobbyist who's ALREADY SUBSCRIBING, so that's not my problem and I don't care how, but solve it. And if you want to punish people for wanting to do something - while you yourself are being greedy - then... maybe it's time to say good bye to any cash that was going to you, and use a local LLM instead?

[hong22333]

不知道有什么办法可以解决？
升级成pro+会不会缓解这种限制
还是说，再订阅一个pro账号轮换使用？

[LegionGPOS]

you would be wasting your money and making this a legit reason for them to keep going. don't give them any more money if you are already on a pro+ account...you literally can't go any higher or change teh rate limits regardless.

你这样做只会浪费钱，而且会让他们以此为借口继续骚扰你。如果你已经是专业版+账户，就别再给他们钱了……无论如何，你都无法升级到更高等级或更改费率限制。

[LegionGPOS]

lol, if you are on a pro account...go ahead...that might give you an extra day.

[Delsin-Yu]

不知道有什么办法可以解决？

No, I enountered this in my Pro+ subscriptions, so there is no way.

[pepperymilkcap]

Has anybody had any responses to any support tickets yet?

[Delsin-Yu]

This outcome was exactly what I expected: GitHub’s level of subsidization appears to have reached a point that is no longer sustainable. As a result, this feels less like a bug and more like a business decision. I’m also skeptical that a request-based commercial model will remain viable in the long run. It’s surprising how many tokens can be consumed in a single conversation turn, and it seems they ultimately chose to shut this down. Still, enforcing both a turn-based request limit and a usage-based request limit at the same time is not user-friendly. I wish they would switch entirely to usage-based limits and provide a visible gauge, rather than locking users out for hours.

[Delsin-Yu]

The 1500 premium request for Pro+ users has become a joke since this change; they are now blocking users based on whatever non-transparent internal AI-slop logic they use.

[LegionGPOS]

Well, what can I say?

Hi there
 
I understand how frustrating it can be to hit the user_weekly_rate_limit. I’d like to clarify that this behavior is intentional and part of ongoing efforts to protect overall service reliability.
 
As GitHub Copilot continues to grow rapidly, we’re seeing increased patterns of high concurrency and intensive usage. While these can come from legitimate workflows, this type of usage place significant strain on shared infrastructure and resources.
 
To maintain a fast and reliable experience for all users, the team has introduced updated limits to better balance system capacity. These changes have now been rolled out and are aimed at ensuring consistent performance and stability across the platform.
 
Users were notified about these updates last Friday. You can find more details in the official announcement here:

https://github.blog/changelog/2026-04-10-enforcing-new-limits-and-retiring-opus-4-6-fast-from-copilot-pro/
 
When you hit a service reliability limit, you’ll need to wait until your session resets. The reset timing should be indicated in the error message when the rate limit is reached.
 
We also recognize the need for better visibility, and the team is actively working on improvements to the user experience, so users can more easily understand and anticipate when they’re approaching these limits.
 
Best,
GitHub Support

exactly, they suck.

[slikeytre]

They need to do what cell companies had to do with unlimited data and just give us slower TPS or find another way to degrade the service without it completely breaking our workflow. Doesn't help that everyone resets at the same time meaning servers get slammed harder on the first of the month, seems like they are making it harder on themselves. What burns the bridge is hiding the Rate limit metric, no reason we shouldn't have a progress bar right now to know when were getting close so we can walk away and take a break for an hour. Lastly the message saying "Reset" at this time is misleading... We know its like a charged ultimate and it just outputs the time where it'll let you request something not that it can actually do anything as you'll just get rate limited again, at least give the true reset time.

[slikeytre]

"you’ll need to wait until your session resets. The reset timing should be indicated in the error message when the rate limit is reached." They act like the service becomes usable if you wait til that time like it's an actual reset.... how disconnected can support be.

[sgaabdu4]

Why is GPT 4.1 rate limited too? I can't use any models. This is crazy. I'm on Pro+

[Delsin-Yu]

It seems that the rule might be that when you hit a rate limit, all options for model selection are inaccessible, leaving only Auto available.

[pepperymilkcap]

The fact that auto is still available makes me suspicious that this is an attempt to curb attempted use of Copilot for specific model distillation.

[LegionGPOS]

I said the same thing, the only other options were Grok and Gemini (or lesser gpt's)...so what is happening exactly!?! are they using lesser reasoning on the models and swapping between them on the fly during conversations? If so, them my 22llm setup is bound to become more useful if I can do that myself. So far its only got governance and memory down I can't figure out the tool calling and fluid conversation just yet.

[eimkasp]

Yep , looks like rate limit is on account level, not per-model basis. Some data and information inside the dashboard to understand this better would be great. People use it for work, so having clarity is very important in my opinion. From work tool, it might become a toy like many other products in the space.

[Zero3K20]

Auto doesn't work for me.

[Noogear]

If this keeps up, I'm literally not renewing next month.

[pepperymilkcap]

It appears that this won't be rectified, so where next? What other services offer similar pricing but actually let us pay and use the service?

[pepperymilkcap]

Perhaps the only thing we can do is wait for the infrastructure to perfect globally.

[LegionGPOS]

Does hugging face offer these kinds of services?

[pepperymilkcap]

I think they do but I cannot speak to the quality. Perhaps it would be worth cancelling Copilot and saving the money for a while until you can buy used GPUs? Then you can run some decent local models (which are getting better all the time).

[LegionGPOS]

I'm super close. I already have open claw activated and setup for what i have going on and i just discovered in need to actuallly create subagents but I already have 80 that are now reactivated properly with python and rust. i just found the ollama extension and set that up...so maybe the timing wasn't bad at all. i'll just abuse the auto since it hasn't stopped yet and i have my steamdeck and main system buzzing.

[UncleMonky]

in what world does anyone need 80 sub agents?

[blazengreen360]

Server Error: Sorry, you've exceeded your weekly rate limit. Please review our Terms of Service. Error Code: user_weekly_rate_limited

Its just Tuesday!

[RandyNorthrup]

I canceled my Pro+ subscription beacause of this 33 hours limited WTF!!

[LegionGPOS]

lol, one of my session said 110

[LegionGPOS]

So they're definitely switching between those "rate'limited" models fyi.

[Delsin-Yu]

Last night, the Auto was routing me to GPT 5.3-Codex, and now it's routing me to GPT 5.4. I'm not sure what the internal logic is for Auto.

[Delsin-Yu]

Smells like an LLM is doing the classification job.

[LegionGPOS]

It tries to use Claude as a sub agent and switches between gpt and Claude still, lowest being 5.3 so I agree. They're leaning and they don't want to admit it or pay up because from all the responses, folks are getting shorted prematurely too. I usually have at least 4 systems with several session running so I expect it but not this harsh. Auto coupled with what I already created seems to sustain somewhat but yeah...they're definitely about lose customers...I think ill just wait until they actually try to charge me before I decide...I could just run it up until they actually close the account.

[chriswhitingBCH]

Also hit the weekly limit after 1 day, but nothing is working in VS chat, auto or specific model - just hangs on the three dots.

[LegionGPOS]

Yeah that happens when I use android studio. I have one system also that can't use anything but gpt (still can't figure out why) so that one's basically useless unless I use anti gravity or my Claude subscription.

[zhw990217]

I also encountered this problem. I encountered a rate limit when subscribing to Pro before, so I upgraded to Pro+. However, today while continuing to work on the project, I was directly informed Sorry, you've hit a rate limit that restricts the number of Copilot model requests you can make within a specific time period. Please try again on Apr 20 at 8 AM. Please review our Terms of Service ( https://docs.github.com/site-policy/github-terms/github-terms-of-service ). (Request ID: 84B2:2B0EDA: 150FC41:17B234F: 69DF48DD). Now only the "auto" mode is available, but I only used less than 20% of the credit limit

[ColdeZhang]

same #192880

[KlotzJesse]

Sources? Sounds AI generated

[wzh5550123]

To the GitHub Copilot Team,

I am writing to express my strong dissatisfaction with your recent restrictions. Are you suggesting that a legitimate individual user running Copilot via remote control between remote and local environments is somehow "violating" your terms?

This is an incredibly overbearing and unreasonable policy. As a paying subscriber, I expect the service to be available across my development environments, whether I am working locally or accessing my setup remotely. Restricting usage based on the method of access—especially when the intent is legitimate personal development—is a blatant disregard for user experience and a breach of the trust we place in a professional tool.

Your current approach feels less like "managing service quality" and more like a forced restriction of a service I have already paid for. I demand a clear explanation for these "intelligence throttling" and access-limiting behaviors.

[pepperymilkcap]

Aaaaand I'm rate limited again.

Still no response to the support ticket that has been open 4 days. Perhaps it would be worth contacting a lawyer?

Edit:

Waited only 30 minutes, tried GPT 5.4 and it worked for a few minutes, then gave the 50 hour rate limit again. What is going on?

[pepperymilkcap]

@Akash1134 Why hasn't GitHub put out a statement on this? Why are paying Pro+ users still getting these extreme rate limits?

[ehendrix23]

Yeah, just reviewing code isn't even working anymore. Few iterations and limited. As of right now GHCP is useless for any more in-depth codebase analysis.

[KlotzJesse]

Copilot CLI

● Changelog for v1.0.32

• Select auto as your model to let Copilot automatically pick the best available model for each session
• Add --print-debug-info flag to display version, terminal capabilities, and environment variables
• Show warnings when approaching 75% and 90% of your weekly usage limit
• Attach supported document files to prompts for the agent to read and reason about
• Add --connect flag to directly connect to a remote session by ID
• /feedback saves the bundle to TEMP when the working directory is not writable
• copilot login --host now correctly authenticates with GitHub Enterprise Cloud (GHE) instances
• Current date and time in agent context now includes local timezone offset
• Terminal progress indicator stays visible while the agent is thinking
• Status line no longer shows stray Unicode glyphs in terminals like Neovim after /clear
• Rewind works correctly after using /cd to change directories
• Multiline input is preserved when using /plan and plan mode
• Backspace correctly exits shell mode only when the input is empty
• Mouse wheel scrolling works correctly in the /ask dialog
• Allow short session ID prefixes (7+ hex chars) with --resume and /resume instead of the full ID
• Rate-limited sessions now pause queued messages and automatically retry instead of dropping them
• Tables render with correct column widths, emoji support, and stable borders during terminal resize
• Rate limit error messages now show specific context based on the type of limit reached
• Session idle timeout is now configurable via --session-idle-timeout; disabled by default
• Skills that exceed the token limit are still discoverable and invocable by name

  Error:
  You've reached your weekly rate limit. Please wait for your limit to reset on April 20, 2026 at 2:00 AM or switch to auto mod
 to continue. Learn More (https://docs.github.com/en/copilot/concepts/rate-limits). (Request ID: 
F24F:35D86:1386FD8:181ED23:69E31DCE) 

Auto in Copilot CLI works under /model

[pepperymilkcap]

The problem is that these "weekly usage limits" shouldn't exist in the first place, because what's the point of paying for a service and paying for premium credits if I'm not allowed to use them?

[CNTianQi233]

Same here.
What makes this especially frustrating is that the UI gives a tiny cooldown like "retry in 3 seconds," but after waiting, the model often just rate limits again. That makes the message feel misleading, because it suggests the issue is temporary when in practice it can keep happening.
I first saw this in a very large long-running engineering session, but later I also saw similar rate limiting in newly created sessions. So this doesn't look like just "your chat got too big" — it looks like something broader is going on.
For paid users, this is a pretty bad experience. If there is a real session, account, or model-level limit, GitHub should be much clearer about it instead of showing a tiny retry timer that doesn't actually resolve the problem.

[CNTianQi233]

GitHub Support replied, but the response was mostly a restatement of the docs.

I understand that global rate limits exist and that premium models have capacity constraints. That part is clear.

What is still not clear is:

• whether the limit is session-level, model-level, account-level, or time-window-based
• whether the “retry in 3 seconds” message reflects a real cooldown at all

That short retry timer is the most frustrating part, because in practice waiting does not reliably solve anything. If the real reset window is much longer, the UI should say so instead of showing a tiny retry countdown that often changes nothing.

I also saw the problem not only in one huge session, but in newly created sessions too. So this does not look like a simple “your chat got too long” case.

At this point the issue is not just the rate limit itself — it is the lack of a clear explanation of how it actually works.

[pepperymilkcap]

Still experiencing this. The worst part is that I'm convinced that "Auto" mode not only selects worse models but configures them at the lowest possible effort setting. I'm having to ask it to continue after every tiny little step it makes, then it condenses and forgets what it's doing.

[Brokenchin]

Hah! I cannot even use Auto or the included models after hitting the limit. I got Pro+. Having a hard time to use all of my included 1500 premium request at this rate. Last month I went way over this and paid the extra out of my own pocket.

I am actually fine with limits. In case they are clearly shown via the plugins themselves. Like weekly 0/100% and maybe shorter 5 hour period. Claude shows me clearly my limit (unofficial plugin for Jetbrains). this allows to plan my work accordingly. Are they harsh.. Absolutely! Using Opus eats a lot of usage out of the 5 hour window. Using Sonnet not so much.

Will be going for refund though from github copilot as these artificial limits should not have happened. Not to mention with these limits there is no way to use all thats included in the price. It's bad business practise. Which seems to be the trend now.

Suddenly applying these limits while not against their TOS was a bad move as a huge business. They should have just eaten the cost and implemented the fixes first. To look good in eyes of their customer. But hey its Microslop nowadays. Should expect as much.

Edit: Not trying to change subject either just wanted to show my disappointment on how Github handled this. I rarely write my thoughts. treat it as rant of an old man.

[LegionGPOS]

One thing i have noticed, depending on your ide and configuration, auto and some models won't do a damn thing. I have one system thats iffy about using autopilot and don't use any other model outside of gpt5.3 and .4. Haven't figured it out and I reinstalled and did a bunch of stuff. I'm using vs code. I know android studio won't use auto at all, I havent tried my copies of visual studio.

Edit: not to change the subject just thought I'd share that as a troubleshooting scenario.

[software-programmer]

I got the 48hr rate limiting yesterday...

Switching to Auto, I've not been limited since, and it's opting for GPT 5.4...if I choose GPT 5.4 instead of auto, I get the rate limit message again.

However, I've still got 75 premium requests on my Pro+

yet I'm being billed for over usage

[pepperymilkcap]

Still experiencing this. 5 days, no reply to my support ticket. Perhaps it would be worth setting up a website?

[CyberExploiter]

Same here. Been a week now

[moonlune]

1. If you have the Pro plan, don’t upgrade to Pro+, because Pro+ still has the same limitations!
2. Given the current instability of the service and the uncertainty surrounding its business model, I don’t recommend subscribing for a full year at once!
3. Some people believe this recent price increase for Token is due to increased demand caused by heavy usage from programs like OpenClaw, but I’m not sure if this is just a marketing ploy or something else entirely!

[Delsin-Yu]

I recently conducted some experiments that support my hypotheses. Using the copilot with this information in mind can significantly reduce the chances of being rate-limited—at least, I no longer experience rate limits by following these rules.

• Tokens/Premium Requests are crucial. The overall "ban" doesn’t trigger based on how much inference you use, but rather on how much inference you consume per premium request.
• Time-based Cooldown Gauge: I believe there is a hidden gauge for each account. When your tok/prem-req exceeds a certain value (possibly defined per model), this gauge heats up, and it cools down after a set period. This hypothesis could explain why users often get rate-limited again when they start using Copilot immediately after the cooldown period ends.

In summary, these rules suggest:

• Each premium request now has less value.
• Overusing a single premium request leads to faster rate limiting, while underusing it makes your premium requests consume more quickly, yet you avoid rate limiting.

With these rules in mind, the ideal usage of the Copilot becomes clear:

• Break down large tasks into multiple rounds of isolated prompt requests.
• Use the auto feature for simpler tasks. I frequently received GPT 5.3 Codex responses when I described the task as complex, and obtained Haiku responses when I didn't specify enough.
• After completing a lengthy task involving multiple sub-agent calls, switch to using auto or perform less token-intensive (short) tasks to balance out the tok/prem-req values.

[LegionGPOS]

This is actually perfect and your skills are crazy but from what I have experienced and based on this...you actually make sense. I've noticed my deeper more thoughtful sessions stayed on sonnet but anything like a tiny fix or a git sync issue was given haiku or lesser gtp's. I never really saw opus usage. So I may need to introduce real discipline here. I wonder how they are rating codex and the others.

[Delsin-Yu]

I've noticed my deeper more thoughtful sessions stayed on sonnet but anything like a tiny fix or a git sync issue was given haiku or lesser gtp's

That's true; the model the Auto mode routes your requests to is equally vague on my side. My colleagues' computers often get routed to GPT 4.1 (which is crazy), and on my work and home computers, I'm experiencing the behavior I described.

I never really saw opus usage.

Based on their latest blog, GPT-5.4, GPT-5.3-Codex, Sonnet 4.6, and Haiku 4.5 are currently the valid candidates.

Auto is dynamic, giving you reliable access to your favorite models while mitigating rate limits. It routes to models like GPT-5.4, GPT-5.3-Codex, Sonnet 4.6, and Haiku 4.5 based on your plan and policies. The models auto will route to will change over time.

This blog kinda invents an entirely new meaning for the English word reliable; it's shocking how fast human language can evolve these days.

[Delsin-Yu]

And THAT's more like it, they are truly cutting costs.

• New signups paused for Pro, Pro+, and Student
• Tighter usage limits for individual plans
• Opus models removed from Pro, and Opus 4.5 and 4.6 are now gone for Pro+ as well
• Reach out to GitHub support between April 20 and May 20 for a refund.

So act fast if you wish for a refund.

Anyway, sounds logical for commercial (aka not losing money), terrible for us customers though.

[Delsin-Yu]

even though the ui in vscode contradicts everything said here

That Upgrade show whenever the GitHub server returns false 😅 , this can happend when you are in an unsupported area, or in Pro, or the GitHub server blows up.

i was curious about 4.7 but i guess that's the stronger of the 3 so they went with that instead.

From people's benchmark, the 4.7 is very poor in consistency, but great in benchmark numbers, sort of Gemini vibes, I think one reasonable explanation counld be that, it's easier to explain to the public by retiring the two low-priced ones in favor of one expensive version (and advertising it to be stronger) comparing to raise the price for the 4.5 and 4.6, this also allow them to concentrate more GPUs on a single unified one, effectively allow more room for more user in the future.

[LegionGPOS]

Definitely a more logical explanation and believable version of the explanation. And yeah touche on vs vode and its inconsistency that was more of a "why can't they get it together" kind of thing for me. Anyway, I haven't really had a chance to sample too much of opus since they've been back up. After seeing how auto works i leave some of the stations on auto and work that I want to be more involved in will use sonnet and 5.4 for a round table kind of interaction.

[pepperymilkcap]

Back to the cheapest QWEN models I can find via OpenRouter I suppose.

[Delsin-Yu]

I highly recommend giving DeepSeek V4 Flash a try. I was surprised by how well it performed on my coding tasks; I would say it rivals both the V4 Pro variant and Sonnet 4.6. Plus, it’s extremely affordable. I've consistently achieved around a 90%+ cache hit rate, and the cost for their cache hits is effectively negligible.

You multiply the price with 0.15 to get the value in dollars.

[Delsin-Yu]

In the future I may just keep a low-level copilot sub (or Cursor sub) for Tab completion and occasionally Claude & GPT access, and spin V4 Flash forever.

We have provided more details about changes to Copilot Individual Plans in this discussion, which may be updated if new information becomes available.

To keep conversations all in one place and to ensure you receive timely updates, we recommend you subscribe to the discussion linked above.