Hey, the ~13K tokens for a “Hi” message is actually expected baseline behavior, not a bug.
Every request to the AI includes built-in overhead that can’t be removed:
- Cursor’s internal system prompt, which tells the AI how to behave
- Built-in tool definitions like file editing, terminal, search, and more, even if you don’t use MCP
- Any active rules, even if they’re minimal
For reference, other AI coding assistants have similar baselines. Claude Code CLI uses about 20K tokens just for “hello,” so 13K is actually on the lower end.
A couple of important things about token costs:
- Most of this is cache read tokens, which are 10x cheaper than regular input tokens
- In the same chat session, later messages reuse the cached context, so the overhead doesn’t grow linearly
To reduce overall usage:
- Keep chats short and focused on a single task
- Start new chats often instead of extending long conversations
- Remove any unnecessary rules or MCP servers
- Keep an eye on which model you’re using since some use more tokens than others
About hitting 84% in 2 days, if you’re using Agent mode a lot, especially with more expensive models, usage can add up fast. You can check the breakdown on your usage dashboard at https://cursor.com/settings to see which models and features are using the most.
There’s a longer discussion here: Saying ‘hello’ uses 122,000 tokens – the cache usage seems inefficient
Let me know if you have more questions.