Mistral's new voice stack supports all 24 EU languages and transcribes from $0.0035 per minute
Mistral launched Voxtral 3 on October 20, 2026: speech recognition, speech synthesis and an SDK for voice agents in a single package, delivered via Mistral's EU-hosted API or as self-hosting. The pricing is aggressive — the full pipeline costs $0.04 per minute according to the company, roughly half of OpenAI's Realtime API. But all the headline figures are Mistral's own, with no independent verification, and several key disclosures are still missing from the launch material.
What was launched
Voxtral 3 consists of three parts connected in a single pipeline: the transcription model Voxtral Small Transcribe 3, a reasoning model such as Mistral Small 4, and the speech synthesis model Voxtral Mini TTS 3. For developers who previously had to set up these links separately, the main message is that the whole chain — listen, think, talk — is now delivered as one streaming service.
The technical foundation is an SDK built on LiveKit's open source infrastructure. According to an MSN-syndicated report on the launch (via MSN, source c4030c80), the SDK sequences all three steps: the transcription model captures and transcribes speech, the reasoning model processes the transcript and can call external tools, and the TTS model converts the response into audio — all streamed as one pipeline. The SDK supports telephony via SIP and browsers and apps via WebRTC (Lapaas Voice, source 4665be65).
Mistral claims under 500 ms realtime latency for transcription and under one second turn latency for the full agent. These are the company's own figures, not measurements confirmed by third parties.
Language support and coverage
Transcription supports 50 languages according to Mistral, including all 24 official EU languages plus several Asian and Indian languages (Lapaas Voice, source 4665be65). Speech synthesis covers a different and smaller number: over 35 languages and over 50 dialects (AI Weekly, source 99c2b956). The two figures should not be conflated — they describe different models.
One detail worth noting for non-European markets: Japanese is supported for both transcription and speech synthesis, but Japanese is not among the pre-made TTS voices, and Japanese speaker-diarization accuracy has not been published (note.com, source 9dd8fa9f).
The numbers — and who stands behind them
Mistral's most eye-catching benchmark is the word error rate on FLEURS 45. According to the company's own figures, as reported by AI Weekly (source 99c2b956), Voxtral Small Transcribe 3 achieves 4.1% when the language is given and 4.0% when it is inferred automatically — versus 16.0% and 14.9% for OpenAI's Whisper large-v3. That is a large difference, but it originates from Mistral itself. The launch post publishes no per-language breakdowns of the FLEURS results, making it impossible to assess how the model performs on individual languages.
For diarization — which speaker said what, and when — the MSN-syndicated report (via MSN, source c4030c80) gives 18.4% cpWER, a metric that combines transcription accuracy and speaker attribution. The Japanese technology blog note.com (source 9dd8fa9f), on the other hand, reports Mistral-measured diarization error for the Mini model of 18.6%, down from 30.9% in the previous version. The 0.2 percentage-point gap may be because the sources describe slightly different measurements, but it is unresolved in the available material.
note.com also states that transcription error for the Mini on Artificial Analysis' index fell from 3.59% to 2.22% compared with its predecessor. The same source, however, rebuts part of the marketing: the claim that Voxtral "beats Whisper and others" is Mistral's own statement, and on the same Artificial Analysis index at least two third-party models have lower error rates than the larger Small model.
The pricing — layer by layer, with one ambiguity
The pricing structure, as reported in the coverage:
- Realtime transcription: $0.008 per minute, with sub-500 ms latency (AI Weekly, source 99c2b956)
- Batch transcription: from $0.0035 per minute for the two batch variants (same source)
- Speech synthesis: $13 per million characters for Voxtral Speech 3 (same source)
- The full agent pipeline: $0.04 per minute, described as roughly half of OpenAI's Realtime API (MSN, source c4030c80)
In addition, Cryptobriefing (source bfe29b5b) reports on a feature called Pro Voice, which is said to make it possible to adapt a highly realistic voice from 30 minutes of reference audio, with an awaited API price of around $0.001 per minute. This figure does not obviously reconcile with the transcription prices, and it is unclear whether it describes a separate pricing tier or something other than per-minute audio. It should be read with caution until clarified.
Sovereign deployment as a selling point
Beyond price, the deployment model is a central point: Mistral offers either its EU-hosted API or self-hosting on private infrastructure, marketed as "sovereign deployment," where the customer controls how long audio is stored (AI Weekly, source 99c2b956). For European organizations building voice interfaces and having to comply with the EU AI Act, control over the data path and audio retention is a concrete selection criterion, not just marketing.
The launch follows Mistral's earlier push into voice: the company launched its first speech understanding models in July 2025 and expanded with TTS in March 2026 (Cryptobriefing, source bfe29b5b).
What cannot be verified yet
Before standardizing on Voxtral 3, there are several gaps in the public material:
- All benchmark figures are Mistral's own. No independent party has confirmed the FLEURS results, the latencies, or the diarization figures. Cryptobriefing and note.com both point to the lack of independent verification, and note.com notes that two third-party models beat the Small model on the same index.
- No customer names. The launch material names no organizations using the stack in production (AI Weekly, source 99c2b956).
- Weights and license are unclear. Although Mistral has historically shipped open weights, neither a weights release nor a license has been announced for this version (note.com, source 9dd8fa9f). This directly affects the value of the "self-hosting" argument: you cannot self-host what has not been released.
- Per-language performance has not been published, for either transcription or diarization.
What this means
The concrete takeaway from the launch is not primarily that Mistral claims records, but that the full voice-agent pipeline is now packaged as one service with one bill, EU hosting or self-hosting, and prices that — according to the coverage — sit below OpenAI's comparable offering. For developers building telephony- or app-based voice agents in Europe, that provides a real alternative to evaluate.
But the evaluation should wait for what is missing: independent measurements, per-language figures, weights with a license, and reference customers. For now, the most usable numbers in the launch are the prices — which are easy to verify by using the API — while the most striking quality claims remain, for the time being, on Mistral's own account.