Mistral AI has released Voxtral 3, a new family of speech-to-text and text-to-speech models paired with a voice agent SDK that combines listening, reasoning, tool calling and speaking into one real-time streaming pipeline. Announced on October 20, the launch is the Paris-based lab’s biggest push yet into voice AI for businesses, and every piece is available immediately through Mistral’s API or for self-hosting on customer infrastructure.

What Was Announced?

Rather than a single model, Voxtral 3 is a complete stack with three parts: Voxtral Small Transcribe 3 for speech-to-text, Voxtral Mini TTS 3 for text-to-speech, and a voice agent SDK built on LiveKit that ties them together with a reasoning model into an agent that can hold a live conversation, reach into business systems and act while the caller is still on the line.

Mistral’s pitch is modularity. “Each stage is a separate model, so you can change the LLM, the voice, or the language without rewriting the agent,” the company says in its announcement. The whole stack is designed for what the company calls sovereign deployment: run it on Mistral’s EU-hosted API or on your own servers, and keep control over where audio is processed and whether any of it is retained.

Transcribe: 50+ Languages, Speaker Diarization, Sub-500ms Realtime

The transcription side is led by Voxtral Small Transcribe 3, which covers 50+ languages, including all 24 official EU languages plus low-resource Asian and Indic languages. It identifies who said what (diarization), accepts key terms and free-form text to bias recognition of domain vocabulary without fine-tuning (context biasing), and detects when a speaker has finished so agents respond promptly (end-of-turn detection).

Mistral claims state-of-the-art results in diarization and multilingual transcription, saying the model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. That comparison comes from the company’s own published tests, not independent evaluation, so treat it as a vendor claim. Independent signal exists anyway: Artificial Analysis has already placed Voxtral Small Transcribe 3 at number three on its speech-to-text leaderboard.

The smaller Voxtral Mini Transcribe 3 packs 4 billion parameters and costs $0.0035 per minute for batch transcription. The realtime variant, Voxtral Mini Transcribe Realtime 3, targets agent use cases with sub-500ms latency in over 35 languages, including diarization and context biasing, at $0.008 per minute. Voxtral Small Transcribe 3 costs $0.0040 per minute.

Voxtral 3 voice agent pipeline diagram
Voxtral 3’s voice agent pipeline: speech in, reasoning, tool calls and speech out run as one streaming system. Image: Mistral AI

Speech: Emotional, Multilingual, and Pro Voice Cloning

On the output side, Voxtral Mini TTS 3 (sold as Voxtral Speech 3) generates speech in 35+ languages and 50+ dialects, adjusting emotion, pacing and emphasis to fit the conversation. It costs $13 per 1 million characters, which Mistral says undercuts comparable models from OpenAI, Deepgram, ElevenLabs Flash and Gemini Flash.

For professional use cases, a Pro Voice feature reproduces a speaker’s distinctive character, emotional range and speaking style across languages from 30 minutes of reference audio, according to the company’s launch details.

The Voice Agent SDK: Phone and Web Calls With Tool Use

The SDK runs the full loop as one streaming pipeline: Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished speaking, a general model like Mistral Small 4 reasons and calls tools, and Voxtral Speech 3 replies in the tone of the conversation so far. Each turn completes in under one second, and the agent stops talking when the caller cuts in (barge-in).

During a call, an agent can verify a caller’s identity, check a bill or book an appointment using function calling, web search and retrieval over the customer’s own systems. It answers and places calls over SIP, runs in browsers and apps over WebRTC, and connects to the telephony, IVR and contact-center software companies already use, as The Decoder’s coverage details.

Pricing and Availability

Voxtral 3 is available now through the Mistral API on European, US and global endpoints, plus Studio playgrounds for testing voice agents, transcription, speech generation and voice cloning. The voice agent API itself is priced at $0.04 per minute when Mistral Small 4 handles reasoning and $0.08 per minute with any other reasoning model, with both rates covering the full conversational workflow. Running the full stack on-premises or in a private cloud is arranged directly with Mistral’s team.

Voxtral 3 launch pricing table
Voxtral 3 launch pricing from Mistral’s official announcement. Graphic: GPTOnline, using Mistral AI assets

Why It Matters

Voice agents for contact centers and service desks are becoming one of the most crowded corners of applied AI, with ElevenLabs, Deepgram, OpenAI’s realtime models and others all competing. Mistral’s entry stands out on three fronts: a modular architecture that lets developers swap any stage without rebuilding the agent, a European-hosted option for data-sensitive customers, and aggressive pricing that starts well under a cent per minute for transcription. Combined with the company’s strong European enterprise position, Voxtral 3 turns Mistral from a model supplier into a direct competitor in the voice-agent platform business.

One caveat applies to the launch, flagged by reviewers of the official announcement: the benchmark comparisons against Whisper, Scribe v2 and Gemini 3.5 Transcribe are Mistral’s own numbers, and pricing is only published for the Mini transcription models, the TTS model and the agent API. Independent evaluations in the coming weeks will show whether the performance claims hold up outside the vendor’s tests.

Sources