When I check my model usage locally it will show like 100k But when I check the usage dashboard it will be 30-40k tokens more.
Is this known? I guess I like to use the token counter to know when the model will enter the “dumb zone” so if there is actually more tokens added at request time (which I assume there are cursors own prompts) then could we get them shown also so we know when the model is creeping above 60% window size?
This is expected behavior. The IDE context indicator and the usage dashboard measure different things:
The context indicator in the IDE shows how full the model’s context window is for your current prompt. It’s an estimate of prompt size relative to the model’s context limit.
The usage dashboard includes all provider-reported token categories: input, output, cache read, and cache write tokens. The 30-40k difference you’re seeing is mostly due to cache tokens, which are a normal part of prompt caching.
So the extra tokens aren’t hidden prompts being silently added. They’re cache read/write tokens that the provider reports for billing purposes, but aren’t part of the “how full is my context window” calculation.
On your second point about surfacing more token detail in the IDE so you can tell when the model is approaching context limits, that’s a fair request. We’ll be tracking this post to gauge interest from the community.