Wide uncertainty by design
An unmeasured route carries deliberately large uncertainty, which pushes its lower-confidence bound down. It therefore cannot satisfy a high quality floor on optimism alone — the router escalates or refuses instead.
Every price carries a source. Every quality tier carries a source. A model with no price on file is infeasible under a cost ceiling — never free.
| Model | Tier | Context | Measured behaviour |
|---|---|---|---|
| claude-opus-5 | premium | 200K | Thinking by default. Measured p90 output is materially higher than non-thinking peers, which the planner accounts for. |
| claude-sonnet-5 | balanced | 200K | Scored 32/32 on our benchmark at 2.6× less than Opus — the strongest single-model baseline we test against. |
| claude-haiku-4-5 | economy | 200K | Failed exactly 5 of 32 cases, all in the hard band: multi-step arithmetic and multi-hop reasoning. |
Anthropic models are available today. OpenAI adapters are implemented but ship without pricing rows, so those endpoints fail closed under a cost ceiling until a price with a recorded source is added.
An unmeasured route carries deliberately large uncertainty, which pushes its lower-confidence bound down. It therefore cannot satisfy a high quality floor on optimism alone — the router escalates or refuses instead.
Graded trials update a posterior seeded by the tier. If measurements contradict the vendor’s claim, the measurements win — otherwise the tier would be unfalsifiable.
Thinking models bill reasoning as output. We use measured output tokens per model and task type, and refuse a candidate whose measured p90 exceeds your token budget rather than returning an empty answer.
The registry is yours to extend — with a price source recorded against each.
Get started