The eval cost two cents. The day was the budget.

544 legal questions through OpenAI's Decisions API, one call each, cost 0.019 dollars: 190,399 input tokens at ten cents per million.

Ten cents per million is the price I typed into my script. The run used about 350 input tokens per question. There was no output-token charge; the endpoint returns a decision, not text. The whole thing ran in about four minutes at eight calls in flight.

The arithmetic holds: 190,399 tokens at ten cents per million is 0.019. What it multiplies is the weak part. The rate was never read off a bill. It matches what OpenAI's Decisions API docs said on 8 October: ten cents per million input tokens, and no charge for output tokens. So 0.019 is an estimate that agrees with the published rate, not an invoice.

The hours went elsewhere: picking six tasks, deciding that a refusal is not a no, rewriting the hearsay prompt after it scored 1 of 9 on a pilot, writing it up so someone could re-run it. None of that has a line in the cost figure.

Limits. The 0.019 is script arithmetic on input tokens at a price I typed in; no invoice, and retried calls are not counted. The four minutes is my stopwatch, not a log. One run, one model, gpt-6-luna.

If an eval costs two cents in tokens and a day in judgement, which of the two should the next budget request be written around?

Where the numbers live

• judge.py
• README.md
• calibrate.py
• FINDINGS.md
• results/metrics.json

The quiz and the cost line in the repo

Script arithmetic, not an invoice. Retries not counted. Four minutes by stopwatch. gpt-6-luna, 8 Oct 2026.