Interface AssemblyAILlmOptions

Gateway model id — see ASSEMBLYAI_GATEWAY_MODELS for the catalog, which is generated from the gateway's own /v1/models and records which models can stream, call tools, and serve the EU region.

Typed against that catalog so a name the gateway does not carry is caught where it is written, rather than as a 400 at the first session. A plain string is still accepted, because the catalog is a snapshot of a service that adds models faster than this package releases.

Note two listed models (gpt-oss-20b, gpt-oss-120b) cannot stream, so they cannot drive a voice pipeline at all.

Defaults to ASSEMBLYAI_LLM_DEFAULT_MODEL.

Reasoning effort forwarded to the model as reasoning_effort.

Unset, no reasoning_effort parameter is sent at all — the model runs on its own server-side default. Set "none" (gpt-5.1 and later) or "minimal" (the original gpt-5/-mini/-nano) to turn reasoning off, e.g. when a voice turn's time-to-first-token matters more than thinking depth. Only GPT-5-family models accept the parameter.

Gateway region. "eu" routes through the EU endpoint for data residency — six models at time of writing, per the eu flag in ASSEMBLYAI_GATEWAY_MODELS. Defaults to "us".