Mistral introduced Voxtral 3 on October 20, 2026, with separate transcription and speech models plus an SDK that connects them to a reasoning model and tools. Its appeal for developers is a documented route from a live microphone to a spoken answer and a business-system action. The release also has practical boundaries: phone access is enterprise-led, model claims need testing in the intended language and setting, and different parts of the stack carry different prices.

A voice agent is a chain of separate parts

In Mistral's voice-agent documentation, the application captures audio. A realtime speech-to-text model transcribes it and detects the end of a turn. A language model writes an answer or calls a tool, and a text-to-speech model plays the response. The SDK manages the conversation and interruptions. The reasoning model is a separate choice, with mistral-small-latest listed as the default in the configuration guide. Each stage is a distinct model or application component.

The distinction matters for a service agent. The spoken request can be transcribed correctly while a tool call still fails, or the model can choose the right tool while producing an awkward reply. Mistral's tool guide splits actions between functions run by the developer's application and tools run on Mistral's side, such as a document library or workspace connector. That placement determines which actions remain available on each channel.

Mistral says its realtime transcription model offers end-of-turn detection, speaker diarization and context biasing, and claims latency below 500 milliseconds. Its release page claims the wider transcription family covers more than 50 languages, while its realtime section says more than 35. Speech generation is advertised across more than 35 languages and over 50 dialects. BIG CHANGE has not measured these figures. Mistral labels both Voxtral Mini Transcribe Realtime 3 and Voxtral Mini TTS 3 Public Preview on their model cards. The current transcription guide distinguishes realtime transcription from file transcription and says the former has no start or end timestamps on its segments. Its realtime API guide documents a WebSocket session, audio-chunk events, incremental text and configurable streaming delay.

A concrete build path: a browser support agent

For a team evaluating a browser-based support agent, the TypeScript browser guide supplies a specific path. A server holding the Mistral API key creates a voice session and returns limited session credentials. The page asks for microphone access, connects to that session, streams audio and receives conversation events. The page then asks the server to close the session. The guide requires HTTPS or localhost for microphone access and warns that the session-creation route needs the application's own authentication. A product team can use this path to establish whether a saved agent hears a request, gives an appropriate answer and calls a permitted support tool. Those outcomes must be checked against real requests and the team's own access controls; the docs do not establish them for any particular deployment.

For a phone line, the architecture changes. Mistral's SIP guide says telephony is available to enterprise customers through an account team, rather than as self-serve setup. The developer brings a number and SIP system. Client-side functions do not run during the call, so required actions must use server-side tools. The same guide currently says calls are inbound only and cannot be transferred to a person. Those limits qualify the launch announcement's broader claim that agents can answer and place calls.

What it costs, and what remains to prove

Mistral's October 20 price table lists Voxtral Mini Transcribe Realtime 3 at $0.008 per minute and the batch Mini transcription model at $0.0035 per minute. It lists Voxtral Small Transcribe 3 at $0.004 per minute, speech generation at $13 per million characters, and a voice-agent API at $0.04 per minute with Mistral Small 4 for reasoning or $0.08 per minute with a larger reasoning model. These are distinct usage modes; teams should price the actual agent configuration and call pattern rather than add every row as if each were always billed together. The company says API service is available through European, US and global endpoints, and directs customers seeking private-cloud or on-premises deployment to its sales team.

The speech model also supports one-off and saved voice cloning. Mistral recommends 30 minutes to three hours of clean audio for professional cloning. A voice recording's technical suitability does not establish the right to use that person's voice. Teams considering this feature need permission, identity and abuse controls in their own product design; the cited guide does not show that a cloned voice is unrestricted or safe in every use.

The documentation has rough edges at launch. An older realtime-transcription URL redirects to a page marked deprecated and naming a Voxtral 2 model; the current Voxtral 3 realtime guide names the new model. Mistral's Voxtral TTS 3 model card says nine languages, while its release announcement advertises more than 35 languages and 50 dialects. Developers need to confirm the exact model, voice and feature combination for their intended locale. BIG CHANGE has not run the SDK or benchmarked its speech quality, latency or task completion.

The big change

Mistral has put transcription, reasoning, tool calls and speech into a documented realtime agent workflow with replaceable stages. Builders now have a specific integration route to evaluate, with separate choices about system access, language coverage, voice and channel. A deployment decision still depends on measured accuracy, task completion and end-to-end delay in its intended setting.

Sources & further reading