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 | In / 1M | Out / 1M | Measured behaviour |
|---|---|---|---|---|---|
| claude-opus-5 | premium | 1M | $5.00 | $25.00 | Thinking by default. Measured p90 output is materially higher than non-thinking peers, which the planner accounts for. |
| claude-sonnet-5 | balanced | 1M | $2.00 | $10.00 | 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 | $1.00 | $5.00 | Failed exactly 5 of 32 cases, all in the hard band: multi-step arithmetic and multi-hop reasoning. |
| deepseek-v4-flash | no vendor tier | 1M | $0.44 | $1.32 | 99.4% over 320 graded trials. Peak-hour price shown; DeepSeek bills half this off-peak, so cost here is deliberately the conservative figure. |
| deepseek-v4-pro | no vendor tier | 1M | $1.32 | $3.96 | 99.7% over 320 graded trials, at roughly 2.7× the cost of Flash for no measured quality gain on this suite. |
| gpt-5.4 | no vendor tier | 1.05M | $2.50 | $15.00 | |
| gpt-5.4-mini | no vendor tier | 400K | $0.75 | $4.50 | |
| gemini-3.5-flash | no vendor tier | 1M | $1.50 | $9.00 | |
| gemini-3.7-flash | no vendor tier | 1M | $0.75 | $3.75 | |
| gemini-3-flash-preview | no vendor tier | 1M | $0.50 | $3.00 | |
| gemini-3.5-flash-lite | no vendor tier | 1M | $0.30 | $2.50 | |
| gemini-3.1-flash-lite | no vendor tier | 1M | $0.25 | $1.50 | |
| glm-5.3 | no vendor tier | 1M | $1.40 | $4.40 | |
| glm-5 | no vendor tier | 200K | $1.00 | $3.20 | |
| glm-4.6 | no vendor tier | 200K | $0.60 | $2.20 | |
| glm-4.5-air | no vendor tier | 128K | $0.20 | $1.10 | |
| glm-5.3-flash | no vendor tier | 1M | $0.075 | $0.25 | Currently listed at a 50% promotional discount, recorded as such in the pricing row's source. |
| gpt-oss:120b | no vendor tier | 128K | unpriced | — | |
| gpt-oss:20b | no vendor tier | 128K | unpriced | — |
Registry snapshot as of 4 September 2026: 19 endpoints across six providers, listing only endpoints that can actually serve a request. Prices are vendor list prices in USD, each recorded against a source in the registry. The two Ollama Cloud models carry no pricing row, so under a cost ceiling they are refused rather than treated as free — that is the behaviour, not an omission from this table. Measured pass rates, latency and cost-per-success for every model listed here — including the ones that score worse than their price suggests — are on the Research page, which is where those figures live rather than being copied into two tables that would drift apart.
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