There is a severe technical bug in how the Dashboard logs usage/requests versus what is actually occurring in the IDE. The usage reporting logic appears to be multiplying actual token counts by >30x in the dashboard logs.
Steps to Reproduce
Open Cursor IDE and make a request using GPT 5.2 xhigh.
Note the "Context Usage" displayed in the chat input box (e.g., 100k).
Wait for the response to finish (or disconnect).
Go to the Cursor Dashboard -> Activity/Usage tab.
Compare the logged token count for that request against the IDE context count.
Expected Behavior
The Dashboard logs should accurately reflect the context passed to the model. A request showing 100k context in the IDE should appear as ~100k in the logs, not 3.4 Million.
For AI issues: add Request ID with privacy disabled
b7d68599-7284-48c8-8469-8ddf2d59ac0b
Additional Information
This has affected me less often for requests in the past few weeks (GPT-5.1) but now it is very consistent the last 3 days. The request ID attached is just one of many.
For example, used up almost half my monthly usage in a SINGLE day because most of the requests suffered from that issue. Ultra plan is useless because if I continue using Cursor as normal, my usage limit will be hit almost immediately.
The two metrics are ultimately measuring different things.
The context indicator in the IDE: Shows how full the model’s context window is for the current prompt. It’s meant to help you see when you’re approaching the model’s context limit.
Dashboard Activity: Shows total accumulated tokens across all model calls in a request. In Agent mode, a single user request triggers many model calls — each tool call (file reads, searches, edits, etc.) is a separate model invocation that re-sends the conversation context. All of these tokens accumulate for billing.
Looking at your specific request (b7d68599....):
Your request triggered 41 separate model calls (agent iterations)
Each call consumed between 29k–100k prompt tokens (avg. 79k tokens)
Total accumulated: ~3 million prompt tokens
I think you’ve mentioned in another post that there aren’t enough tool calls to justify this level of token use – if you can share a Request ID with privacy mode turned off I’d be happy to take a look!
Hi @Colin Thanks a lot for your response. I understand better now the difference between total usage and context window in IDE. However, the usage is still far beyond what I would expect. I went ahead and turned privacy mode off as suggested, and initiated a new follow-up request in my conversation from the last request ID I sent.
This time it is 8M tokens which is pretty absurd - that’s probably multiple times more than my entire codebase tokenized. Context usage in IDE shows 189.2K. Notably the majority of the tool calls seem to just be reading a few hundred lines a file or running greps. It is quite strange if each tool call is using tens of thousands (or hundreds of thousands, to explain the 8 million) tokens.
Request ID (made with privacy mode disabled):
be3c75e6-abe4-44d6-8981-1f39681ac27d
@Colin Just following up: I did some more testing on a completely different codebase yesterday (one that is significantly smaller and would completely fit in 100-200k tokens) and also faced the same issue. The tool calls for reading a tiny portion of a specific file or running a grep is simply impossible to add up to the number that’s billed. I also see it with the GPT 5.2 high (not extra high) variant.
I also got an email response which dismissed my issue as just MAX mode pricing. So I want to make it clear: MAX mode was not enabled for any of these requests.
It seems to be from tool calls and/or applying edits (which I guess is also a tool call technically) in both Agent mode and manual mode. The issue doesn’t seem to occur when using a custom mode with zero tools enabled. Something is not right with tool call usage.
Here is another request that’s pretty much definitive proof: Chat context was ~100K before the request, I made the request using a custom Manual mode. The ONLY tool calls were file searches (which all returned 100% empty, so obviously not token-heavy) and file edits (of just <100 lines of code in total). Counted as 2.8M tokens of usage!
Thanks a lot for those Request IDs. Indeed, something doesn’t look quite right. I’ve filed a bug report, and I believe the team is already aware of another case.
