In the “Agent Window” mode, you can choose to limit tokens to 256k, but the context window will grow to 500k (in the case of grok 4.7). In the case of opus 5.5, a window of 300k is selected, but the context grows to 1M. I’ve used up 40% of my API limits because of this…
Steps to Reproduce
Activate the context window in 256k on grok 4.7 and start filling it
You’re right: with Opus 5.5 set to 300K, the chat should summarize before approaching 1M. I confirmed that your Opus 5.5 chats were not being held to 300K. We’ve also reproduced the same behavior with Opus 5. This is a known issue we’re tracking, and we’ve shared your report with the team.
For Grok 4.7, your 256K selection was respected, and those chats stayed under 256K. If you see Grok exceed that limit, please share the Request ID and a screenshot of Context Usage.
For now, the safest workaround is to monitor Context Usage and start a new chat before approaching 300K.
Did the context actually grow beyond the limit selected in the UI?
Which requests/tokens were actually responsible for the quota consumption?
The screenshots + Request ID are probably the most important evidence for Cursor to check the first one server-side.
For the second one, if you still have the Usage export, I’d preserve it and isolate the events around the affected runs: model, timestamp, input/output tokens, cache-read/cache-write and any unusually large requests. That can at least show whether the quota spike lines up with the sessions where the context apparently exceeded the selected limit.
I wouldn’t assume that “500k context shown = 500k newly billed tokens”, because cached context and Cursor’s internal accounting can make those numbers different. But if the selected cap was 256k and the actual context kept growing well past it, the Request ID gives Cursor something concrete to investigate.
I built a small local CSV analyzer for this kind of Cursor usage reconciliation, if it’s useful as a second check:
Disclosure: I built the analyzer. It won’t prove Cursor’s server-side cause, but it can help separate what the Usage export actually shows from what we’re inferring from the UI.
Hey @Maksim_Ermakov, following up: the Opus 5.5 issue is now fixed. With 300K selected, the chat now runs against a 300K limit and summarizes before it reaches it.
The fix was applied on our side, so no update is needed, and any chat that’s already past 300K will summarize on your next message. If you see it go over the limit again, please share a Context Usage screenshot and the Request ID here.
One more small update on the Grok 4.7 part of your report: the Context Usage panel was showing a 500K limit even with 256K selected. That display has been corrected on our side too, so with 256K selected it now reads against 256K.
Thanks again for the report, and let me know if you see either model go past the limit you picked.
