The Pulse
Mistral Releases Voxtral 3 Voice-Agent Stack
Mistral released Voxtral 3 on October 20, 2026, pairing speech models with a LiveKit-based SDK for voice agents that can transcribe, reason, call tools and speak. The models support transcription in more than 50 languages, and Mistral lists
AI.info Team · Written with AI · fact-checked ·
Under one second per conversational turn is Mistral’s stated target for Voxtral 3, a voice-agent stack the company released on October 20, 2026. The system combines speech recognition, a reasoning model and speech generation in a streaming pipeline, with an SDK for connecting agents to business tools and phone systems.
Mistral is selling the stack as a way to build agents that can do more than transcribe and answer: they can retrieve information, call functions and complete tasks while a conversation is underway. The company’s examples include checking a bill, verifying a caller’s identity and booking an appointment. Those are product capabilities described by Mistral, not independent performance results.
Three stages run together, not one after another
The SDK, built on LiveKit, coordinates separate models for listening, reasoning and speaking. Voxtral Mini Transcribe Realtime 3 detects when a caller has finished; a general-purpose model, such as Mistral Small, handles reasoning and tool calls; Voxtral Speech 3 generates the spoken reply. Mistral says the stages run concurrently rather than waiting for each other, which is how it arrives at its under-one-second turn-time claim.
“No stage waits for the one before it, so each turn completes in under one second, and the agent stops talking when the caller cuts in.”
Mistral
The SDK supports calls over SIP and browser or app sessions over WebRTC, and is designed to connect with existing telephony, interactive voice response and contact-center software. Because each stage is a separate model, developers can change the language model or voice without rewriting the agent.
Transcription spans 50 languages; speech generation spans 35
Mistral says Voxtral transcription supports more than 50 languages, including 24 official European Union languages. The transcription models also offer diarization, which identifies who spoke when, and context biasing: developers can provide names or technical terms to improve recognition of domain-specific vocabulary.
The realtime transcription model is listed at sub-500-millisecond latency. Mistral’s launch page describes its realtime language coverage as both over 35 and, elsewhere, over 40 languages; the company gives the broader 50-plus-language figure for its transcription family. Voxtral Speech 3 supports more than 35 languages and 50 dialects, with controls for emotion, pace and emphasis. Mistral also says its voice adaptation can use 30 minutes of sample audio.
Transcription starts at $0.0035 per minute
Mistral lists batch Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Small Transcribe 3 at $0.004 per minute, and the realtime Mini transcription model at $0.008 per minute. Voxtral Speech 3 costs $13 per million characters for batch and realtime use. The full agent stack can be accessed through Mistral’s API; the company says on-premises or private-cloud deployment is available by contacting its team.
The release also makes a data-location claim: Mistral says organizations can use its European-hosted API or deploy the stack on their own infrastructure to control where audio is processed and whether conversations are retained. Its API is listed across European, US and global endpoints, so deployment choices depend on the endpoint and setup an organization selects.
Mistral’s accuracy claims come with a company-run comparison
Mistral says its Small transcription model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription and diarization. The launch page presents benchmark results, including evaluations on public datasets, but the comparison is published by Mistral itself. Its speech-generation comparisons also rely on tests described on the company’s page, including evaluations judged by Gemini 3.8 Flash.
For businesses, Voxtral 3’s main proposition is a single route from spoken request to action, with the option to swap components and select an API region or private deployment. Its published prices cover the speech models; the launch page does not give a separate price for the voice-agent SDK or a bundled rate for an entire call.