The inputs
- Stable prefix: 2,500 tokens. System prompt and tool schemas.
- New tokens each step: 600. Tool results and new instructions.
- Output each step: 350. Whatever you want billed as output, including reasoning if you have that count.
- 8 steps. Cache on, 5-minute writes. Batch off. 100 runs a day.
What the transcript does
Step 1 sends 3,100 tokens. Step 8 sends 9,750. Across the run the prompts add up to 51,400 input tokens, which is 16.6 times the first step, not 8 times. Of those, 9,750 are cache writes and 41,650 are cache reads. Output is 2,800 tokens.
The bill for this shape
- $0.0607 this run.
- $6.07 a day at 100 runs.
- $182.11 over a 30-day span.
Turn the cache off and this same run costs about $0.0701 more. One step with the cache on costs more than one step with it off, because you pay the write and never read it back. The receipt says so when that happens.
Open this example in the calculator
Change the token counts. These ones are not a benchmark, and the total is not an invoice. Retries, hidden reasoning, and a prefix too short to cache will not match.