Routing

Models & routing

Code routes by slot, not model — the registry resolves a slot to a concrete model per call and records it on every run. Multi-provider by design; the judge slot never falls back (one eval label = one model).

Slots

judgeno fallback

Per-criterion LLM-as-judge scoring (generateObject, one call per criterion)

Default

claude-opus-4-7

eval integrity — one label, one model

reportfallback

Narrative report generation over scored results

Default

claude-sonnet-4-6

gemini-2.5-pro → gemini-2.5-flash

utilityfallback

Cheap structured helper tasks (classification, extraction, titles)

Default

gemini-2.5-flash

claude-haiku-4-5

Model catalog

ModelProviderContextIn /1MOut /1MCapabilities
claude-opus-4-7anthropic200K$15.00$75.00
streamingtoolsstructured-outputvisionprompt-cache
claude-sonnet-4-6anthropic200K$3.00$15.00
streamingtoolsstructured-outputvisionprompt-cache
claude-haiku-4-5anthropic200K$1.00$5.00
streamingtoolsstructured-outputvisionprompt-cache
gemini-2.5-progoogle1M$1.25$10.00
streamingtoolsstructured-outputvisionthinking
gemini-2.5-flashgoogle1M$0.30$2.50
streamingtoolsstructured-outputvisionthinking
gpt-5.5-proopenai1M$30.00$180.00
streamingtoolsstructured-outputvisionprompt-cache
gpt-5.5openai1M$5.00$30.00
streamingtoolsstructured-outputvisionprompt-cache
gpt-5.4-miniopenai400K$0.75$4.50
streamingtoolsstructured-outputvisionprompt-cache
gpt-5.4-nanoopenai400K$0.20$1.25
streamingtoolsstructured-outputvisionprompt-cache