AI Release Gate

Cost

From the per-call records, through the DuckDB read model. Each run's total below is checked against the figure its RUN.json recorded.

By run and model configuration

Run	Model configuration	Calls	Errors	Input tokens	Output tokens	Cost
2026-09	`anthropic-alias`	2,100	0	986,140	359,404	US$2.78
2026-09	`anthropic-snapshot`	2,100	0	986,140	360,059	US$2.79
2026-09	`anthropic-sonnet-snapshot`	2,100	0	1,300,700	323,121	US$5.83
2026-09	`google-alias`	2,100	2	776,246	490,251	US$2.43
2026-09	`google-snapshot`	2,100	1	792,454	472,653	US$2.38
2026-09	`openai-alias`	2,100	0	433,780	146,924	US$1.02
2026-09	`openai-snapshot`	2,100	0	447,604	148,144	US$1.04
2026-09	`openweights-control`	2,100	2	968,352	247,192	US$1.26
2026-09-run2	`anthropic-alias`	2,100	0	986,140	360,881	US$2.79
2026-09-run2	`anthropic-snapshot`	2,100	0	986,140	361,897	US$2.80
2026-09-run2	`anthropic-sonnet-snapshot`	2,100	0	1,300,700	330,483	US$5.91
2026-09-run2	`google-alias`	2,100	2	802,161	479,119	US$2.41
2026-09-run2	`google-snapshot`	2,100	0	798,153	484,049	US$2.42
2026-09-run2	`openai-alias`	2,100	0	406,132	145,610	US$1.00
2026-09-run2	`openai-snapshot`	2,100	0	391,796	146,027	US$0.99
2026-09-run2	`openweights-control`	2,100	6	968,139	248,849	US$1.27

Run totals

Run	From the calls	Recorded in RUN.json
2026-09	US$19.53	US$19.53
2026-09-run2	US$19.57	US$19.57

By block, all runs

Block	Calls	Cost	Mean output tokens	Largest output
long_context_recall	1,600	US$11.95	12	195
closed_form_reasoning	9,600	US$9.09	191	4,096
instruction_following	4,800	US$6.63	272	4,136
multiple_choice	8,000	US$4.15	87	1,656
refusal_calibration	3,200	US$3.43	199	1,961
paraphrase_robustness	3,200	US$2.52	145	640
structured_extraction	3,200	US$1.34	44	166

Read-only. Built 2026-09-28T17:59:46Z from commit 02a709fa02cf of the repository, where every number here can be regenerated offline with the CLI. Intervals are 95%.