Mistral launches Voxtral 3: a voice agent SDK and new speech models

Mistral AI announced Voxtral 3 on 20 October 2026, describing it as "a new family of state-of-the-art speech-to-text and text-to-speech models, alongside a voice agent SDK". Mistral's documentation lists each of the new models as "Public Preview".

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude. Mistral sells language models that compete with Claude, and its charts compare Voxtral with OpenAI and Google products.

What's in the release

Mistral's pricing table lists:

• Voxtral Small Transcribe 3, which the docs call Mistral's "flagship transcription and diarization model": $0.0040/min.
• Voxtral Mini Transcribe 3: the post says "The Mini variant is 4B parameters"; $0.0035/min for batch.
• Voxtral Mini Transcribe Realtime 3, which Mistral says "achieves high-quality transcriptions at sub-500ms latency": $0.008/min.
• Voxtral Speech 3, the text-to-speech model (also called Voxtral Mini TTS 3 in the post and Voxtral TTS 3 in the docs): $13/1M chars.
• Voice agent API: "0.04$/min" with Mistral Small 4 for reasoning, "0.08$/min" with any other reasoning model.

For transcription, Mistral says "The supported list of languages is expanded to 50, including 24 official EU languages". The counts vary: the post says the realtime model works "in over 35 languages" in one place and "across 40+ languages" in another, and gives the speech model "35+ languages and 50+ dialects", while the speech model's docs page says it "Supports 9 languages".

The voice agent SDK

Mistral says Voxtral 3 "introduces a voice agent SDK, built on LiveKit, that runs speech in, reasoning, and speech out as one streaming pipeline." Mistral says "each turn completes in under one second".

The post says the agent "answers and places calls over SIP, runs in browsers and apps over WebRTC", and that "Each stage is a separate model, so you can change the LLM, the voice, or the language without rewriting the agent".

Where it runs, and what about weights

Mistral says "Voxtral 3 is available today through the Mistral API, hosted through our European, US, and global endpoints." Elsewhere the post refers to "Mistral's EU-hosted API or for self-deployment on your own infrastructure", and its pricing section says: "To run the full stack on-premises or in a private cloud, talk to our team."

The post does not mention downloadable weights or a licence. When we checked Hugging Face on 21 October, the newest Voxtral repository under Mistral's account was Voxtral-Mini-4B-Realtime-Arabic, created on 8 October; there was no Voxtral 3 repository, so we are not calling these models open-weight. Our mid-2026 open-weight scoreboard noted "the Voxtral TTS audio model (March)".

The benchmarks, in full

Mistral publishes nine charts. We read every row.

• FLEURS, 45 languages (7 models): Voxtral Small Transcribe 3 and Microsoft AI Transcribe 2 both score 4.0% word error rate with the language specified and 4.1% with it auto-detected. Voxtral Mini Transcribe 3 scores 4.5%/4.6%, Gemini 3.5 Transcribe 4.5%/6.1%, Scribe v2 4.9%/4.9%, GPT Transcribe 5.6%/6.3% and Whisper large-v3 14.9%/16.0%.
• Diarization (7 models): Voxtral Small Transcribe 3 leads at 18.4%, then Voxtral Mini Transcribe 3 (18.6%), Pyannote Orchestrator (Precision 3) at 19.0%, Microsoft AI Transcribe 2 (20.3%), Scribe v2 (23.0%), Voxtral Mini Transcribe 2 (30.9%) and Gemini 3.5 Transcribe (38.4%).
• Keyphrase recall (3 models): with context, Voxtral Small Transcribe 3 scores 94.6% to Scribe v2's 94.3%; without context, Scribe v2 leads, 84.0% to 81.4%. Gemini 3.5 Transcribe scores 87.3% with context and 80.9% without.
• Artificial Analysis WER index (37 models): Fun-Realtime-ASR-preview is first at 1.73%, Microsoft AI Transcribe 2 second at 2.04%, Voxtral Small Transcribe 3 third at 2.10% and Voxtral Mini Transcribe 3 fifth at 2.22%.
• Realtime FLEURS, 35 languages (9 entries): Gemini 3.5 Realtime Transcribe scores 4.2%, ahead of all four Voxtral Mini Realtime 3 latency settings (5.7% at 2400 ms to 7.4% at 240 ms) and Microsoft AI Transcribe 2 Streaming (5.8%); GPT Live Transcribe scores 8.4% and Scribe v2 Realtime 8.7%. An offline Voxtral Mini Transcribe 3 baseline scores 4.0%.
• Text-to-speech (4 charts): Voxtral TTS is third of 16 on Voice Code Bench, which Mistral credits to BeSimple AI (80%, behind Eleven v4 and Eleven v4 Turbo at 98%); first of 10 plotted on pronunciation control (87%); second of 10 plotted on multilingual quality (3.16, behind Eleven v4 at 3.17); and third of 16 on expressive English (3.13, behind Eleven v4 at 3.17 and Cartesia Sonic 3.6 at 3.15). Three of these charts are labelled "Internal eval", and all four name gemini-3.8-flash as the judge.

Mistral says Voxtral Speech 3 "costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash". Its own price axis lists Qwen-Audio 3.1 Flash at $2, and Gemini 3.8 Flash Lite TTS and Speechify Simba 3.2 at $10, all below Voxtral's $13; the other twelve rivals plotted are priced between $15 and $85.

Why it matters

On our reading, Voxtral 3 is mainly a packaging move: transcription, reasoning and speech sold as one per-minute voice agent, with European hosting and self-deployment by arrangement. Mistral's own charts show it leading some tests and trailing others. On our reading, with the models in public preview, prices and language lists may still change.