I switched to ‘Max’ mode for processing a large set of markdown files with Opus within 1M context window.
The files total size was 1.2 Mb, far less than 1M, but nevertheless it started ‘compacting’ the context in the middle of the very first reply, and even after I stopped it (because compaction destroys my goal entirely!) the context status ‘circle’ was still showing 120% full over 300K size!
It nevertheless charge me $4.6, although did NOT produced any useful output at all!
These are two clear bugs in the design!
Steps to Reproduce
Start new session, feed 1.2 Mb of markdown files, instruct to read them and produce something.
Expected Behavior
tokens accounting system switches to 1M context length IMMEDIATELY after I change to ‘MAX’ mode and checkbox 1M context
allow user to selectively: a) disable, b) trigger compaction (all other agentic environments already allow that!)
Thanks for the report, and sorry you ran into this!
Unfortunately, I haven’t been able to reproduce the issue on my end. As soon as I select the 1M variant of a model, Max Mode is automatically enabled, and summarization doesn’t kick in until most of that context window has been used. Everything in the backend tied to your Request ID indicates that you should be seeing the same behavior.
Would you mind sharing a couple of screenshots to help us dig deeper? Specifically:
A fresh chat configured with a 1M context window (any dummy prompt is fine), showing your open model selector.
Here’s how it looked on my end in contrast (this screenshot was made immediately after the ‘compaction’ started in a situation I described in the first post) - attached image.
The attached video shows how, in a agentic chat session, the context metrics don’t get updated after I change model.
Regarding the video you shared: the context usage indicator only updates after a new message is sent. This is expected behavior.
That said, I was able to reproduce the issue you described. When Max Mode is enabled, the larger context window is not automatically selected. (The reverse does work: selecting the larger context window automatically enables Max Mode.)
Looking into this on the backend, I found that when Max Mode is enabled, requests are actually sent using the larger context window configuration, even if 300k is the selected option. This mismatch leads to unexpected behavior, such as conversations containing more tokens than should be permitted.
As a workaround, you can manually select the larger context window, which should prevent early summarization. I’ll pass this along to the team for further investigation.
Thank you for deep investigation and confirming the issue!
Could it be also possible to provision some reasonable payment-free quota for 1M context models on my account so that I could safely ensure the workaround works?
Additionally, I discovered that with Gemini 3.5 Flash, for which even with Max mode turned off, Cursor UI shows ‘1M context’, it actually treats it as 200k context model, and starts compaction!
Please fix that, and please refund in some form what I’ve already lost in experimenting with that!
