Cursor Auto picks a different model every time and my usage ran out in 9 days

The dashboard shows per-request cost and model if you expand a row. Sort by cost; you'll probably see that the expensive ones are long agent runs deep into a chat, not a particular model.

What helped me:

• New chat per task. A 60-message chat resends a lot of context with every request.
• Pin a model for agent work so the cost is predictable, use Auto / a small model for quick questions and Tab.
• Set a spend limit in settings so you get a warning instead of a surprise.

Plus: when the agent is stuck (3rd attempt at the same fix), stop. Every retry in a failing loop costs as much as the first try, and the context gets worse, not better. New chat with a two-line summary of what you know works better and is cheaper.

Pricing and what Auto routes to has changed a few times this year, so check the current docs page rather than old Reddit threads. The advice above about chat length holds either way.