Smoke-run cost estimate for phase one (smoke2)

From 225 smoke calls (5 seed cases x 5 graders x 9 judges, one epoch). Routing: gpt-4o-mini, gpt-4.1 (OpenAI), Opus 5.5 (Anthropic) and Gemini 3.7 Flash (Google) direct; the rest via OpenRouter with require_parameters (gpt-oss-20b also ignores CoreWeave). OpenRouter judges use billed usage.cost; direct judges use Inspect total_cost from llm-prices. Projection: mean cost per call x 6,000 generic calls + mean x 1.5 x 2,400 repo-scorer calls.

judge	cost per call (USD)	mean output tokens	mean reasoning tokens	projected (USD)
claude-opus-5-5	0.00192	32	12	18.46
gpt-4.1	0.00056	24	0	5.42
glm-5.3	0.00054	76	48	5.18
gemini-3.7-flash	0.00043	76	54	4.10
gpt-5-mini	0.00032	137	93	3.05
gpt-3.5-turbo-0613	0.00024	27	0	2.29
gpt-4o-mini	0.00004	26	0	0.42
gpt-oss-20b	0.00002	94	72	0.23
llama-3.1-8b-instruct	0.00002	25	0	0.16

Total projected phase one: $39.31. Smoke2 cost $0.0928.

Seed cases are short; real SimpleQA answers will raise input tokens, so allow 2x.