Gemini 3.8 Live Extended Thinking: the id, and Sume's audio ids

The model id is gemini-3.8-live-extended-thinking, a generally available audio-to-audio model on the Gemini Live API for background reasoning during live audio. Sume does not document a live audio-to-audio route, so there is nothing to swap it for; Sume's audio model ids are for music, text to speech and transcription jobs.

Gemini facts are from the Gemini API changelog. Sume facts are from Music Router and the OpenAPI, read 2026-10-01.

How do the two Gemini 3.8 Live ids differ?

The changelog lists two: gemini-3.8-live, the default for most low-latency voice agent experiences and real-time dialogue, and gemini-3.8-live-extended-thinking, a high-reasoning model that supports background reasoning during live audio. Pick by whether the agent needs the extra reasoning.

What audio model ids does Sume expose?

Sume uses Sume-owned public ids and catalog ids, discoverable from the API.

What happens with an unknown id?

On the Music Router, an unknown id fails with 400 model_not_found and a catalog_url. Omit model to get sume/music-auto; job.request.routed_model names the engine that ran.

Does the job pattern change for audio?

Yes, compared with a live session: Sume audio is submit, poll or webhook, fetch. A synchronous wait is capped at 30 seconds by wait_timeout_seconds; see Jobs and results and the Gemini TTS polling post.