Credits

How the prepaid balance works, what draws it down, and how to avoid running dry mid-run.

Credits are a prepaid balance spent per token. Every request deducts input and output tokens at that model's rate.

What draws the balance down

• Input tokens — the whole request: system prompt, full history, tool schemas, images.
• Output tokens — what the model generates, at 4–5× the input rate.
• Nothing else. There is no per-request fee and no monthly minimum.

Reading your usage

Every response carries a usage object with exact token counts — that is the billing basis, and it is worth logging alongside your own request ids. The dashboard aggregates the same data per model and per key.

Not running out mid-task

• Long agent runs consume credits faster than chat because history is resent every turn.
• A 402 mid-session stops work immediately — top up before a long run rather than during one.
• If usage is steady and heavy, a plan is usually cheaper than credits.