Mistral launches Voxtral 3 voice agents with limits on phone access

A voice-agent toolkit with separate routes to deployment

Mistral released Voxtral 3 on October 20, pairing new speech models with a software development kit for live voice agents. Developers can build conversational services around transcription, a reasoning model, tools and spoken replies. Phone deployment has narrower conditions: Mistral's documentation reserves it for enterprise customers.

The launch lists voice-agent API rates of $0.04 per minute with Mistral Small 4 for reasoning, or $0.08 with a larger model. It advertises European, US and global API endpoints; running the full stack on premises or in a private cloud requires contacting Mistral. Those API rates do not establish a complete telephone-service cost.

Source 1Source 2Source 3Source 6

The SDK takes on the conversation pipeline

The practical addition is coordination. Mistral's overview describes an SDK that captures speech, transcribes it, obtains a model response and plays synthesized audio. It also handles people interrupting the agent. Python and Node.js applications can use the SDK, while the browser implementation uses TypeScript.

Developers can still assemble separate transcription, chat and speech services, but then own the audio transport, delays and turn handling themselves. With the SDK, they instead configure the agent's behavior and integrations. A saved agent holds instructions, model choices, voice and tool declarations; each conversation runs as a separate session.

Saved configurations are versioned, and sessions can pin a version. Our reading is that this lets a support team keep an established configuration in service while trying a revised one.

Source 3

A browser assistant still needs a secure application

Mistral's browser guide requires a backend to create sessions with the API key. The page receives connection credentials and streams audio directly to Mistral. Microphone access needs HTTPS or localhost. Keeping the key off the page is only one boundary: the guide also requires the application's own authentication on the session-creation route and warns against relying on browser tool handlers for authorization.

Tools determine what the assistant can accomplish. Mistral runs built-in search, document retrieval and workspace connectors on its side; connectors requiring authentication must already be authenticated. Custom functions instead execute in the developer's application and return results to the model.

Consider a hypothetical order-status assistant. A custom function could fetch a delivery record from the application's backend, then let the model explain it aloud. Saving that function's declaration does not save its executable handler: the application must attach the handler when opening a session. Without it, the SDK reports a failed tool call. A conversational front end therefore still depends on working business integrations.

Source 4Source 5

Phone support has a narrower launch boundary

The telephony guide says customers bring their phone number and calling system, which forwards calls to Mistral using SIP, a standard for internet calling. Setup goes through Mistral's team rather than a self-service console.

There is also a material documentation conflict. The October 20 announcement describes agents answering and placing calls; the telephony guide inspected on October 21 lists inbound calls only. Outbound availability therefore needs clarification. Separately, the guide explicitly says the agent cannot transfer a call to a person.

Phone sessions cannot execute client-side functions, either. The order-status example would require an integration supported by the phone service; its browser-side custom handler cannot run there. Our assessment is that a browser demonstration does not establish that the same agent can run a complete telephone support workflow.

Source 1Source 4Source 6

A fast first sound is not a finished answer

Mistral's launch advertises sub-second conversational turns. Its latency documentation adds a more useful distinction: model processing time excludes the network and application work, while the caller experiences the entire sequence from finishing a sentence to hearing a reply.

That sequence includes detecting the end of the speaker's turn, transcription, reasoning, any tool call and speech generation. The documentation says tool calls are often the longest wait because they depend on another service. Short spoken fillers can cover the silence while a lookup runs, but do not make the result arrive sooner.

For the hypothetical delivery lookup, an immediate acknowledgment and the actual delivery date are different milestones. Mistral recommends measuring both first sound and useful answer; a fast acknowledgment alone cannot substantiate a fast completed task. We have not measured either.

Source 1Source 7

Existing audio integrations face concrete changes

The Voxtral 2 migration guide identifies changes beyond replacing model names. Word-level timestamps are removed: requesting them returns a 400 error, while segment timestamps remain. Realtime transcription settings must be supplied before the first audio chunk; later updates produce an error event.

Speech generation adds inline emotion tags while continuing to accept earlier preset voices. The guide says the latest TTS alias will move to Voxtral 3 within a month of general availability. Its deprecation-timeline section supplies no retirement date, so it does not establish how long older-model access will continue. Existing integrations need to account for the documented behavior changes without assuming an unstated access window.

Source 8