OpenRouter AI SDK provider

Only for the evaluation path. experimental_evaluate ships in [email protected] and newer, while the rest of the package works with any ai@7. That line requires Node.js 22+ and is ESM-only; older SDK lines are pinned separately at 1.5.4 for v5 and 2.9.1 for v6, so check which major you are on before upgrading.

Treat it as alpha. It is served from OpenRouter's /api/alpha endpoint and the README warns that request and response shapes may change ahead of the rest of the OpenRouter API. It also rounds probabilities and scores to two decimals; the model declares that rounding on every result so the SDK accepts scores that differ from the exact weighted mean by rounding alone.

No. A score question whose criteria contains null, or a boolean question whose criteria describes only one of true and false, throws an InvalidArgumentError before anything is sent and before anything is billed. A boolean question with no criteria descriptions at all is simply sent without criteria.

The AI SDK contract has no slot for them, so they are preserved under result.providerMetadata.openrouter.answers keyed by question id for per-answer confidence and score legend, with the request cost under providerMetadata.openrouter.usage.cost. Boolean questions map to OpenRouter noul questions and surface as probability, while choice and score keep their distributions.

The current release line supports ai@^7.0.0, requires Node.js 22 or newer and is ESM-only. experimental_evaluate itself ships in [email protected] and newer, while the rest of the package works with any ai@7; the legacy 2.9.1 and 1.5.4 lines target AI SDK v6 and v5.

AI SDK boolean becomes an OpenRouter noul question and its noul value is exposed as probability, while choice and score map directly and keep their probability distributions. Per-answer confidence and score legends are preserved under result.providerMetadata.openrouter.answers, and cost is under providerMetadata.openrouter.usage.cost.

It throws an InvalidArgumentError, before sending anything and before anything is billed, for a score question whose criteria contains null and for a boolean question whose criteria describes only one of true and false. A boolean with no criteria descriptions is sent without criteria.