API: pin an agent's TTS model, and Simba 3.5 Turbo

API: pin an agent’s TTS model, and Simba 3.5 Turbo

An agent can now pin the Simba model it synthesizes on, and a new endpoint lists the models available to pin.

`GET /v1/agents/tts-models`

Lists the models a voice agent can be pinned to. The response is the whole catalog, identical for every workspace — which model serves an agent is a routing decision, not an entitlement, so nothing here is plan-gated, and the list is not paginated. Each entry carries the facts that decide whether a given agent may select it:

Field	What it says
`id`	The value to send as `tts.model`.
`display_name`, `description`	Human-readable labels.
`beta`	A training still under evaluation. Fully servable; just not one agents are steered onto by default.
`default`	The model an agent lands on with no pin and an uncurated voice. The only entry selectable for every agent and every voice.
`english_only`	The model has no multilingual deployment, so an agent that resolves through the multilingual serving cannot select it.
`requires_voice_curation`	A voice must be registered for the model before it can be paired with it. The voice’s `models` array on `GET /v1/agents/voices` is the per-voice answer.

`tts.model` on the agent

AgentTTSConfig gains model (nullable). Null is the default and the recommended setting: the agent follows the model its voice is curated for, so a voice promoted to a newer training moves with it and no configuration goes stale. GET /v1/agents/voices now returns default_model per voice — what “automatic” resolves to for that voice. Send null or "" to clear a pin back to automatic.

A pin the agent could not actually be served with returns 400: an English-only model on an agent that resolves through the multilingual serving (it declares additional_languages, or its language is not English), or a model the chosen voice is not curated for. The pair is re-checked whenever voice_id, language, or additional_languages change, so an edit elsewhere cannot orphan a stored pin.

The model pinned at call start is recorded as tts_model on the conversation config snapshot (null = the automatic choice).

Simba 3.5 Turbo

AgentVoiceModel.name adds simba-3.5-turbo — the enum is now simba-3.0, simba-3.2, simba-3.5-turbo. Treat it as an open set and fall through for values you do not recognise. As of 2026-08-27, simba-3.5-turbo is offered on every first-party Speechify voice (its requires_voice_curation is false), except voices built for another model, whose speaker embedding is valid for that training alone; partner voices list no upgrade tier at all.

Nothing you already send changes — an agent with no tts.model runs on its voice’s curated model exactly as before.