Hey, I checked your usage for the last month and there are no anomalies or changes in token counting. The pattern is stable: on active days you’re getting about 1,500 to 3,000 requests via toolCallComposer in Auto mode.
What could make it feel like “it’s more”:
On April 20, the chat was set to claude-4.6-opus-high-thinking. That’s one of the most expensive models, about 28 USD for just 34 requests. Compared to normal days, that’s a noticeable spike.
Agentic sessions, when the agent makes tool calls, reads files, edits, and verifies, can generate dozens to hundreds of separate model requests from a single prompt, and each one counts. That’s by design, but on long autonomous tasks the usage grows fast.
You can see the daily, model, and feature breakdown at cursor.com/dashboard. It’ll show exactly where the tokens are going.
If you want to reduce usage:
For routine tasks, keep Auto, don’t turn on Max Mode unless you need it
Use Thinking or opus models only for the hard tasks
Provide a narrower context and more specific prompts so the agent doesn’t scan files unnecessarily
If you spot a specific day or a specific chat where the numbers really don’t match what you did, send the Request ID and we’ll dig in.
There is a thread about high cache use somewhere. I’ve been working on streamlining rules, skills and tonight cursor ignore file to make sure its not pulling in stuff like .venv etc. to early to see how this works. I have to do something that want to pull the whole codebase.