Token limit exceeded” on GPT-4.1 despite input being far below limit

[Lopat6]

Hi

I'm encountering a "token limit exceeded" error when using GPT-4.1 via Copilot, even though — based on my estimate — the input size is nowhere near the 1 million token limit.

I tested the same input with Gemini 2.5 Pro, which also supports 1 million tokens, and it processed the request without any problems. In both cases, the actual input is far below the stated token limit.

Since I currently have to estimate token usage manually, it would be really helpful if Copilot or the API provided an input identifier or token count feedback so that users can better understand and debug such issues. As a workaround, I’ve been forced to start new chats and reinsert the full context, which is quite inefficient.

Could you please check whether this is a bug or an undocumented limit in how token usage is measured in GPT-4.1? I'm happy to provide sample inputs or more context if helpful.

Thanks in advance!

Best regards,
L.

[Lopat6]

For comparison, I tested the same input with Claude 3.7 Sonnet (which supports up to 200k tokens), and it processed the request without any issues. This suggests that the problem is likely not with the input size itself, but possibly with how GPT-4.1 measures or enforces token limits within Copilot. The discrepancy in behavior between models might help in diagnosing the issue.

L.

[mikejhill]

Running into a similar issue.

We are experimenting with context lengths and techniques to manage within our org. I am noticing both Claude 3.7 and GPT-4.1 erroring out on token limit around the same 200K character length:

• "Edit" mode: "Oops, the token limit exceeded. Try to shorten your prompt or start a new conversion."
• "Ask" mode: "maxTokens must be larger than the ellipsis length"

I have tested for an exact limit and there is some randomness near the boundary on whether the error will be triggered, probably from the response generation length.

https://platform.openai.com/tokenizer counts my input at somewhere around the ~60K token mark (plus a few thousand when including context files), so my best guess is Copilot is limiting requests to 64K tokens, but I'm not confident.

Would be helpful to have both clear documentation on any token limits enforced by Copilot, rather than the underlying models. Would also be helpful to have clear feedback sharing the number of tokens submitted.

Edit:
I see Copilot itself has a 64K-token context, seemingly regardless of the underlying model. Originally mentioned here for 4o and then confirmed by someone (non-authoritatively?) here for 4.1, although I still can't find an official resource to back this up.

[dieuluucanh]

Copilot has hard limit of max prompt size for gpt-4.1 at 64k token. gemini-2.5-pro at 128k tokens. There is no mention of 1 million max context length in copilot's official doc. This limit of context length align with industry standard, like cursor and windsurf. Although cursor has a lot more generous length limit and state it clearly in their docs
Current limit is understandable for base model like Gpt-4.1, however, I really hope github extend base models to 128k or more in future

[sumowesley]

I started receiving "I'm nearing the token limit for this thread" last week. I've spent the last few months using Claude Sonnet 4 and it had been fine. Now I get these warnings within 20 minutes of working, perhaps sooner. It renders agent mode useless. Given that agent mode has the habit of doing more than it needs and keeps every file it has amended/created open, it makes it even more likely that the limit will be reached sooner rather than later. Is this a temporary issue, or should I accept that agent mode is no longer as capable as it once was?

[Aedial]

I got this error in the middle of an agentic task, after several files had been edited. I would have expected the plugin to just compress the conversation or remove unnecessary information from the context, like is done with Claude Code.

[sumowesley]

I got this error in the middle of an agentic task, after several files had been edited. I would have expected the plugin to just compress the conversation or remove unnecessary information from the context, like is done with Claude Code.

The issue has rendered agent mode pretty much useless for anything other than small tasks. I asked it to create some tests yesterday on an additional method I'd added to a DI service: it was in response to that single prompt, referencing three files, that it hit the token limit within a few minutes. It's quite a change to the usefulness of the agent service.

[jiec-msft]

Hey Aedial sumowesley We put some fix in nightly release: 1.5.57.13255-243-nightly

Could you give it a try and let me know if the token limit issue could be mitigated?

[Aedial]

Hey Aedial sumowesley We put some fix in nightly release: 1.5.57.13255-243-nightly

Could you give it a try and let me know if the token limit issue could be mitigated?

I seems to work? Threw a couple of heavy agentic tasks to it and I haven't gotten the message again, but further testing is warranted, I guess.