Hey, good question. The ~18.5k tokens for a “hello” in an empty project is actually expected baseline behavior, this isn’t a bug.
Every agent request includes overhead you can’t fully remove:
- Cursor’s internal system prompt, which tells the AI how to behave
- Built-in tool definitions, like file editing, terminal, search, etc., even if you don’t use MCP
- Any rules, even minimal ones
For reference, other AI coding assistants show similar baseline numbers. Claude Code CLI uses ~20k for a simple “hello”, and Codex in Cursor is about the same. So 18.5k is actually on the lower end.
A few things to keep in mind:
- Most of this is cache read tokens, which are 10x cheaper than regular input tokens
- The overhead is roughly constant per message, it doesn’t scale linearly with more messages in the same chat because caching kicks in
- Later messages in the same chat reuse the cached context
To minimize token usage overall, keep chats short and focused on one task, start fresh chats often, and avoid unnecessary rules or MCP servers.
There’s an active discussion with more details in this thread: Saying ‘hello’ uses 122,000 tokens – the cache usage seems inefficient
Let me know if anything else is unclear.