Mistral has released Voxtral 3, a set of transcription and speech models plus a voice agent SDK that runs listening, thinking and speaking as one stream. The announcement page carries a publish time of 12:00 UTC on October 20, 2026, and Mistral posted the launch thread on X at 16:00 UTC.
The pitch is the whole loop, not one model: “Speech transcription and voice generation have come a long way, but building end-to-end voice agents that can listen, reason, interact, and handle complex workflows remains difficult.” This page reads Mistral’s post directly, prices and charts included.
On this page · 6 sectionsOpen
Mistral released Voxtral 3 on October 20, 2026: three transcription models, a speech model and a voice agent SDK, available through its API on European, US and global endpoints.
Transcription costs $0.0035 a minute for Mini, $0.004 for Small and $0.008 for the realtime model; speech costs $13 per million characters.
Voxtral Small Transcribe 3 covers 50 languages and leads Mistral’s diarization chart at 18.4% cpWER, against 38.4% for Gemini 3.5 Transcribe.
The voice agent SDK, built on LiveKit, streams listening, reasoning and speaking together, and Mistral says each turn completes in under one second.
Every benchmark in the post is Mistral’s own, and its speech evals are judged by another lab’s model, gemini-3.8-flash.
§ 01What Mistral released
| Model | What it does | Price |
|---|---|---|
| Voxtral Small Transcribe 3 | Highest-accuracy transcription, 50 languages, diarization, context biasing | $0.0040 per minute |
| Voxtral Mini Transcribe 3 | Batch transcription, 4B parameters | $0.0035 per minute |
| Voxtral Mini Transcribe Realtime 3 | Live transcription under 500 ms, end-of-turn detection | $0.008 per minute |
| Voxtral Speech 3 | Speech in 35+ languages and 50+ dialects, batch and realtime | $13 per million characters |
Two features matter for anyone building agents. Context biasing lets you hand the transcriber key terms, such as product names, so it spells them right without fine-tuning. Pro Voice, for professional voice cloning, works “using just 30 minutes of audio.” Mistral also names the speech model two ways in the same post, Voxtral Speech 3 and Voxtral Mini TTS 3.
§ 02How the voice agent works
The SDK is “built on LiveKit,” the open-source real-time audio framework. The realtime transcriber listens and spots when the caller stops talking, a general model such as Mistral Small reasons and calls tools, and Voxtral Speech 3 answers. Mistral’s description of the timing: “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.”
It answers and places calls over SIP, runs in browsers and apps over WebRTC, and each stage is a separate model, so you can swap the language model or the voice without rebuilding the agent. The whole stack runs on Mistral’s API or on your own infrastructure.
§ 03What Mistral’s charts show
| Model | cpWER |
|---|---|
| Voxtral Small Transcribe 3 | 18.4% |
| Voxtral Mini Transcribe 3 | 18.6% |
| Pyannote Orchestrator (Precision 3) | 19.0% |
| Microsoft AI Transcribe 2 | 20.3% |
| Scribe v2 | 23.0% |
| Voxtral Mini Transcribe 2 | 30.9% |
| Gemini 3.5 Transcribe | 38.4% |
On plain multilingual transcription across 45 FLEURS languages, Voxtral Small Transcribe 3 scores 4.1% word error with the language given, and Mistral writes: “It outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe.” Its own chart puts Microsoft AI Transcribe 2 level with it at 4.1%, a tie the sentence leaves out.
The realtime chart is the honest one to read. Voxtral Mini Realtime 3 gets more accurate the longer it waits before writing: 7.4% error at 240 ms, 6.4% at 480 ms, 6.0% at 960 ms and 5.7% at 2400 ms. Gemini 3.5 Realtime Transcribe sits at 4.2% on the same chart, ahead at every setting, while GPT Live Transcribe (8.4%) and Scribe v2 Realtime (8.7%) trail.
For speech, Mistral says Voxtral Speech 3 “costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.” Its quality charts plot judge scores against price, and those are internal evals scored by gemini-3.8-flash, another lab’s model, on prompts Mistral selected.
§ 04Where it fits
Voxtral 3 is aimed at contact centers, service desks and hands-free assistants, and its strongest argument is control: an open pipeline you can host yourself, in Europe if you need to, at per-minute prices well below a cent. Next to all-in-one voice models such as Gemini 3.8 Live, it is a parts kit: you choose the brain, and you own the plumbing.
§ 05What we are watching for
- Independent tests of the transcription and diarization numbers on outside data.
- Whether Mistral publishes weights for any Voxtral 3 model, which the post does not mention.
- Real-world latency on phone lines through SIP, where Mistral’s under-one-second figure will be tested.
§ 06Sources
- Mistral, Voxtral 3, October 20, 2026, 12:00 UTC, including the price list and all charts. Read directly.
- Mistral AI on X, launch thread, October 20, 2026, 16:00 UTC.
- Mistral, documentation, for the API and Studio playgrounds the post links to.
- Our records on Gemini 3.8 Live and Gemini 3.8 Flash TTS.
Q1When did Mistral release Voxtral 3?
October 20, 2026. The announcement page carries a publish time of 12:00 UTC, and Mistral’s X thread went up at 16:00 UTC.
Q2Which models are in Voxtral 3?
Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, Voxtral Mini Transcribe Realtime 3 and Voxtral Speech 3, which the post also calls Voxtral Mini TTS 3.
Q3Can I run Voxtral 3 on my own servers?
Mistral says yes: the models run on its API or on your own infrastructure, and the on-premises route goes through its sales team.
Q4How many languages does it support?
Transcription covers 50 languages, including the 24 official EU languages. Speech covers 35+ languages and 50+ dialects. The realtime model is listed at 35 in one place and 40+ in another.
Q5Does CellCog use Voxtral?
No. CellCog’s agent tiers run on Claude Opus 5.5, and its AI employees’ meeting and phone voice runs on another vendor’s live voice model. We follow voice models because they decide how an AI employee sounds on a call.