Mistral Ships Voxtral 3: 4 Speech Models and a Voice Agent SDK at $0.004 a Minute

Mistral AI released Voxtral 3, a family of four speech models plus a voice agent SDK built on LiveKit, with all four models available in public preview on the Mistral API. The release puts the Paris-based company in direct competition with Microsoft AI, ElevenLabs and Google in a speech-to-text market where independent rankings now separate winners from marketing claims.

The family ships three transcription models and one text-to-speech model. Voxtral Small Transcribe 3 serves as the flagship for batch transcription and diarization, with documentation also listing translation, speech question-answering, summarization and function calling. Voxtral Mini Transcribe 3 is a 4B-parameter model for cheaper and faster batch jobs. Voxtral Mini Transcribe Realtime 3 handles streaming ASR over WebSocket with end-of-turn detection. Voxtral Speech 3 covers text-to-speech. Mistral did not disclose the parameter count of Small Transcribe 3 or the GPU and VRAM requirements for self-hosting, though the company says Mini suits resource-constrained environments.

Both batch models list a 32k context window.

How accurate is Voxtral Transcribe 3?

The headline number comes from Artificial Analysis, the independent benchmarking service. Small Transcribe 3 scores 2.1% AA-WER, placing it third on the non-streaming leaderboard. Mini Transcribe 3 scores 2.2%. Only Fun-Realtime-ASR-preview at 1.7% and Microsoft's MAI-Transcribe-2 at 2.0% rank higher.

Mistral's self-reported numbers reinforce the batch strength. On FLEURS 45 with forced language, Small scores 4.1% WER, Mini 4.6%, and Whisper large-v3 16.0%. With inferred language, Small scores 4.0% and Whisper 14.9%. MAI-Transcribe-2 ties Small exactly at 4.1% forced and 4.0% inferred.

Diarization is where Mistral claims the clearest lead. Across five public datasets, Small Transcribe 3 posts 18.4% cpWER and Mini posts 18.6%. The open-source Pyannote system scores 19.0%, and MAI-Transcribe-2 scores 20.3%. In the comparison table Mistral published, ElevenLabs Scribe v2 records 23.0% and Gemini 3.5 Transcribe 38.4%.

All transcription models support diarization and context biasing. Context biasing accepts key terms or free text to improve recall of domain vocabulary. On the Earnings 22 keyphrase benchmark, Small reaches 94.6% recall with context biasing versus 81.4% without. ElevenLabs Scribe v2 posts 94.3% and 84.0% on the same measure.

Batch transcription covers 50 languages, including all 24 official EU languages. That European coverage matters: Mistral sells deployment on European, US and global endpoints, and OpenRouter now lists all three new models with EU-hosted, zero-data-retention endpoints.

Where does Voxtral 3 fall short?

Speed and realtime accuracy are the gaps. On the Artificial Analysis leaderboard, Small Transcribe 3 runs at a speed factor of 65.8x, against 307.5x for MAI-Transcribe-2. Scribe v2 sits at 65.4x and Gemini 3.5 Transcribe at 89.5x. In practice, Mistral's model processes audio slower than the Microsoft alternative it nearly matches on accuracy.

The realtime model trades latency for accuracy on FLEURS 35. At 2,400 ms latency it scores 5.7% WER. At 480 ms the figure rises to 6.4%. At 240 ms, WER reaches 7.4% — well above the 4.2% that Google's Gemini 3.5 Realtime Transcribe achieves. MAI-Transcribe-2 Streaming scores 5.8%. GPT Live Transcribe at 8.4% and Scribe v2 Realtime at 8.7% trail every Voxtral realtime setting.

Mistral announced no open weights, a notable omission for a company that built its reputation on releasing open models. Enterprises wanting to run the models on their own infrastructure must go through Mistral sales.

What does the voice agent SDK do?

The SDK runs three stages as one streaming pipeline on LiveKit. Realtime Transcribe 3 listens and detects when the caller stops speaking. A general LLM, such as Mistral Small 4, reasons and calls tools. Voxtral Speech 3 then replies in the tone of the conversation so far.

Each turn completes in under one second, and the agent stops talking when the caller interrupts. The SDK handles phone calls over SIP and browser or app sessions over WebRTC. Because each stage is a separate model, developers can swap the LLM, the voice or the language independently — an architecture choice that distinguishes the offering from single-model voice systems.

How much does it cost?

Mistral's price table positions the batch models aggressively:

• Voxtral Mini Transcribe 3 (batch): $0.0035/min
• Voxtral Small Transcribe 3: $0.004/min
• Voxtral Mini Transcribe Realtime 3: $0.008/min
• Voxtral Speech 3 (batch and realtime): $13 per 1M characters
• Voice agent API: $0.04/min with Mistral Small 4, $0.08/min with any other reasoning model

By Artificial Analysis pricing, which works out to $4.00 per 1,000 minutes for Small Transcribe 3, the model sits between MAI-Transcribe-2 at $1.67 per 1,000 minutes and Gemini 3.5 Transcribe at $5.00. ElevenLabs Scribe v2 costs $3.67 per 1,000 minutes. Microsoft's vendor list price of $0.10 per hour is a limited-time offer; ElevenLabs lists $0.22 per hour, and Google bills Gemini transcription by token.

What about text-to-speech?

Voxtral Speech 3 costs $13 per 1M characters, a price Mistral says undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. The model page lists roughly 90 ms time-to-first-audio and 34 emotion tags. Explicit controls cover emotion, pacing and emphasis. A feature called Pro Voice adapts a speaker's style from 30 minutes of audio. Artificial Analysis places the model in the top 12 of both of its TTS arenas.

Why does this release matter?

The competitive picture is tight at the top of speech-to-text. Small Transcribe 3, MAI-Transcribe-2 and Scribe v2 all score between 2.0% and 2.2% AA-WER, so buyers will choose on price, speed, diarization quality and data residency rather than raw accuracy alone. Mistral's argument bundles those factors: best-in-chart diarization at 18.4% cpWER, EU-hosted zero-retention endpoints via OpenRouter, and a voice agent pipeline that completes turns in under a second.

Microsoft retains the edge on speed and price; Google retains the edge on realtime accuracy. What Mistral is selling is the integrated loop — listening, reasoning, tool calls and speaking in one API, at $0.04 per minute with its own LLM. If the company closes the realtime gap with Google and loosens its grip on weights, Voxtral 3 becomes the more complete speech stack; until then, it is the strongest batch transcriber with the fastest-growing list of caveats.