Mistral launched Voxtral 3 on October 20, bringing speech models and a voice-agent SDK into one release. The SDK connects transcription to a reasoning model, tools and spoken replies. For developers building a phone service, access has a narrower scope: the hosted route for answering calls requires an enterprise arrangement. The realtime model card dates the release and labels that model a public preview.
This is a documentation-based assessment. We haven't run the SDK or tested a call, so Mistral's speed claims remain unverified here.
A browser prototype needs different tools on the phone
The SDK overview describes a LiveKit-based pipeline with Python and TypeScript bindings. Developers define what the agent knows and which tools it can call, while the SDK handles audio capture, transcription, model responses and playback. Users can interrupt it mid-sentence.
Client-side tools fail on phone calls. The overview explains that a phone session has no application running the handlers for those functions. A browser demo that looks up an order through application code therefore needs a different tool path before it can answer the same question on a phone line. Mistral directs developers to server-side tools for that work in its integration guidance.
Hosted telephony requires enterprise access. The SIP guide says it isn't self-serve. A saved agent answers calls to your number. Customers bring the number and calling system, then contact Mistral to enable the service. That access step belongs in a delivery schedule before a team promises a phone launch.
Measure the wait for an answer
Mistral's launch post claims turns complete in under one second. Its latency documentation defines the useful measurement more precisely: per-turn latency runs from when the user stops speaking to when the agent starts answering. That includes turn detection and the work surrounding the models.
The latency guide also warns that tools often cause the longest wait. A short spoken filler can cover a tool call, but it doesn't shorten the operation. A team evaluating an order lookup should measure time to the useful answer separately from time to the first sound. Otherwise, a prompt acknowledgment could hide a slow backend.
Existing Voxtral integrations need a migration check
The Voxtral 2 migration guide identifies a breaking change: word timestamps are gone, and requesting them returns a 400 error. Segment timestamps remain supported. Realtime settings must also arrive before the first audio chunk, after which the session locks them. Those details affect transcript displays and session setup even if an application never adopts the new agent SDK.
The next useful evidence would be a phone evaluation using the intended server-side tools, with the wait for a substantive answer measured from the caller's end. Mistral's SDK guidance explains why replaying a browser conversation won't establish that the phone integration works. Enterprise access and the tool path need to be settled before that test can begin.