Mistral’s Voxtral 3 Transcribes Speech From $0.0035 a Minute

Mistral AI introduced Voxtral 3 on 20 October 2026: new speech-to-text and text-to-speech models and a voice agent SDK. Its price table lists Voxtral Mini Transcribe 3 at $0.0035 a minute and Voxtral Speech 3 at $13 per 1M characters.

Mistral’s launch thread on X says the SDK is built on LiveKit: “Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation.”

Business Pill · LATENCY VS THROUGHPUT

A short explainer of latency versus throughput. It teaches the general idea only and says nothing about Mistral or its models.

What Mistral Launched

The announcement describes four parts. Voxtral Small Transcribe 3 handles speech-to-text, which Mistral says covers 50 languages, including 24 official EU languages, with diarization (who spoke when) and context biasing for domain terms.

Voxtral Realtime Transcribe 3 is the streaming version. Mistral says it “achieves high-quality transcriptions at sub-500ms latency in over 35 languages,” and that end-of-turn prediction is natively supported.

Voxtral Mini TTS 3.0 generates speech. Mistral says it supports 35+ languages and 50+ dialects, adjusts emotion, pace and emphasis to the conversation, and offers a Pro Voice adaptation built from 30 minutes of a speaker’s audio.

The fourth part is the voice agent SDK. Mistral says the agent can verify a caller’s identity, check a bill or book an appointment during a call, and that it answers and places calls over SIP and runs in browsers and apps over WebRTC.

What Was Posted on X

Mistral AI posted the launch on X at 16:00 UTC on 20 October, introducing Voxtral 3 as transcription and speech models powering its new voice agents API. The prices and benchmark claims in this piece come from the company’s announcement.

In the same thread, Mistral said Voxtral 3 also introduces a voice agent SDK built on LiveKit, in which a transcription model listens, a reasoning model thinks and calls tools, and a speech model replies. These are the company’s descriptions; we have not tested the SDK.

The Prices and Claims Mistral Gives

Mistral’s price table lists Voxtral Mini Transcribe 3 (batch) at $0.0035/min, Voxtral Mini Transcribe Realtime 3 at $0.008/min, Voxtral Small Transcribe 3 at $0.0040/min and Voxtral Speech 3 at $13/1M chars.

On transcription, Mistral says the Small model “outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe,” and that the Mini variant has 4B parameters. On speech, it says: “At $13 per 1M characters, Voxtral Speech 3 costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.”

The announcement says several of its speech comparisons are internal evaluations scored by a model judge, and that prices were taken from vendor pricing pages. These are the company’s figures and comparisons; we have not tested the models.

Built for Sovereign Deployment

Mistral says the Voxtral 3 family is “designed for sovereign deployment.” It says the models are available through the Mistral API on European, US and global endpoints, and that customers who want the full stack on-premises or in a private cloud should talk to its team.

The announcement also says each stage of the agent is a separate model, “so you can change the LLM, the voice, or the language without rewriting the agent.”

Connecting the Coverage

Voxtral 3 arrived on the same day ElevenLabs launched Eleven Realtime, whose v1 Max model, ElevenLabs says, starts speaking in 570ms on average, with an API compatible with OpenAI’s, as we covered in ElevenLabs’ OpenAI-compatible voice model. The two companies quote different measures: Mistral’s sub-500ms figure is for realtime transcription, while ElevenLabs’ 570ms is v1 Max’s average time to first audio, so the numbers do not compare directly.

OpenAI’s changelog of 15 October added a native Decisions integration in GPT-Live, which we covered in our Decisions API piece. Earlier this month Mistral launched a public preview of Mistral Large 4, which it calls a 1 trillion-parameter model, and on 28 September, in the post announcing its Munich hub, it said it would build one gigawatt of European compute capacity by 2030.

This week Traders Union reported that Mistral chief executive Arthur Mensch said it would be a fundamental mistake for Europe to leave frontier AI to American and Chinese companies, as we noted in our update on Germany’s AI pact.

The Structural Read

As we read it, price is part of the pitch: the announcement lists Voxtral Mini Transcribe 3 at $0.0035 a minute and says Voxtral Speech 3 costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.

As we read it, the sovereignty pitch is the European angle: Mistral says the family is designed for sovereign deployment, on its API or on a customer’s own infrastructure.

As we read it, the comparisons are the company’s: the announcement says several of its speech evaluations are internal and that prices were taken from vendor pricing pages.

Three Implications

PRICED BY THE MINUTE Mistral lists transcription from $0.0035 a minute and speech at $13 per 1M characters.

SWAP ANY STAGE Mistral says each stage is a separate model, so the LLM, the voice or the language can change without rewriting the agent.

COMPANY CLAIMS The benchmark comparisons are Mistral’s own, several of them internal evaluations.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework puts it this way: “The surface, not the model, governs autonomy.”

As we read it, Mistral is also selling a surface: its SDK answers and places calls over SIP and runs in browsers and apps over WebRTC, according to the announcement. The framework was published in July 2026 and does not discuss Mistral.

What Is Not Established

The benchmark comparisons are Mistral’s own, several of them internal evaluations, and we have not reviewed the underlying data. The announcement we read does not say whether Voxtral 3 weights will be released, and it gives no price for on-premises deployment.

The Bottom Line

Mistral launched Voxtral 3 on 20 October 2026: transcription models priced from $0.0035 a minute, a speech model at $13 per 1M characters and a voice agent SDK built on LiveKit. It says the family is designed for sovereign deployment. The benchmark claims are the company’s.