Mistral Voxtral 3 Launches With $0.04/Min Voice Agents
The speech release combines transcription, speech generation and a LiveKit-based agent SDK, with separate language limits, explicit pricing and private-deployment options.
Listen
AI narration
12:05
0:00 / 12:05
AI SummaryGenerated from this article
Mistral launched Voxtral 3 on October 21, 2026, a voice-agent platform combining transcription, text-to-speech, and an LLM-powered agent SDK at $0.04 per minute using Mistral Small 4.
The release includes six separately priced products, from $0.0035/min batch transcription to $0.08/min agent configurations with alternative reasoning models. Language support varies by endpoint — 50 languages for file transcription, 45 for real-time, 35 for TTS. Mistral claims sub-second latency and competitive FLEURS accuracy, though these are vendor-reported figures. Private deployment is offered, but publicly downloadable model weights are unconfirmed.
Mistral is offering a voice-agent API at $0.04 per minute with Mistral Small 4, alongside separately priced transcription and speech-generation models. Its Voxtral 3 announcement brings listening, LLM reasoning, tool calling and spoken responses into one real-time pipeline.
The October 21, 2026 announcement covers two file-transcription models, a real-time transcription model, text-to-speech with voice cloning, and a voice-agent SDK built on LiveKit. Organizations can use hosted APIs or pursue deployment on their own infrastructure.
Voxtral 3 gives production voice teams several options, each with qualifications. Language coverage changes by endpoint, the speech documentation contains a conflicting language count, and Mistral’s sub-second latency and comparative accuracy results are vendor-reported. Self-deployment also does not establish that the model weights are publicly downloadable.
The Model Choices and Published Prices
Developers can buy individual speech capabilities or use Mistral’s integrated voice-agent offering. The published rates use two billing units: audio minutes for transcription and agents, and text characters for speech generation.
Product
Model identifier or configuration
Published price
Voxtral Mini Transcribe 3
voxtral-mini-transcribe-3-0
$0.0035 per minute for batch transcription
Voxtral Small Transcribe 3
voxtral-small-transcribe-3-0
$0.0040 per minute for batch transcription
Voxtral Mini real-time transcription
voxtral-mini-transcribe-realtime-3-0
$0.008 per minute
Voxtral Mini TTS 3
voxtral-mini-tts-3-0
$13 per million characters
Voice-agent API with Mistral Small 4
Integrated agent configuration
$0.04 per minute
Voice-agent API with other reasoning models
Alternative reasoning-model configuration
$0.08 per minute
At those rates, 1,000 minutes of batch transcription costs $3.50 with Mini or $4 with Small. Real-time transcription costs $8 for the same duration, while the agent configurations cost $40 or $80.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
With only $0.50 separating the two batch models per 1,000 audio minutes, accuracy on a team’s recordings could matter more than the price difference. Mistral identifies the Mini transcription variant as a 4-billion-parameter model, though that figure alone does not establish self-hosting requirements.
Speech generation needs a different calculation. At $13 per million characters, 100,000 characters cost $1.30. There is no fixed conversion to an audio-minute price because speaking pace, pauses and the supplied text affect output duration.
These are separate product rates, not an instruction to add every row together. Before budgeting an integrated application, buyers should confirm what the agent rate includes, how minutes are metered and whether separately purchased services incur additional charges. Private-deployment costs also need to be checked independently of hosted API pricing.
The announcement says the batch transcription coverage includes all 24 official EU languages, alongside Asian and Indic languages. Both batch models support diarization, which identifies who spoke when, and context biasing, which supplies vocabulary or contextual text to help recognize specialized terms.
Speaker attribution is useful for meeting records and multi-party calls; context biasing targets names, product terminology and other words that a general transcription model might misrecognize. Neither feature proves accuracy on a particular organization’s recordings.
For file transcription, developers can specify a two-letter ISO 639-1 language code or let the model detect the language. Real-time sessions similarly allow a language setting or automatic detection. Speech generation follows the language of the input text, while preset voices have their own language and accent characteristics.
One documentation inconsistency remains unresolved. The supplied Voxtral TTS 3 model card says the model supports nine languages, while the announcement and dedicated language table describe substantially broader coverage.
The dedicated table is the more detailed planning reference, but developers should confirm their target language against the actual endpoint before committing. A general claim of multilingual support is insufficient when transcription, streaming transcription and speech generation have different lists.
The SDK Connects Speech to Business Actions
Mistral’s voice-agent SDK is built on LiveKit and supports tool calling, SIP and WebRTC. Developers can build telephone and browser-based conversations, with an LLM handling reasoning between speech recognition and generated responses.
Tool calling lets an agent request information from a connected system or invoke an action during the conversation. Mistral’s announcement illustrates this with an appointment-booking scenario that retrieves a referral and calls a scheduling tool.
The vendor demonstration does not establish that arbitrary business workflows will work reliably. Developers still need to define permissions, validate tool arguments and decide when an action requires explicit user confirmation. A speech system that correctly hears a request can still take the wrong action if the application’s tools are poorly constrained.
The TTS offering also includes zero-shot voice cloning. Separately, Mistral describes Pro Voice adaptation using 30 minutes of audio to reproduce a speaker’s character, emotional range and speaking style across languages.
That 30-minute figure describes Mistral’s claimed adaptation input; it does not guarantee voice similarity for every speaker. Organizations considering Pro Voice should confirm access conditions and adaptation pricing directly. The listed $13-per-million-character generation rate does not by itself establish the commercial terms for creating a professional adapted voice.
Sub-Second Latency Is a Vendor Claim, Not a Guarantee
Mistral describes under-one-second voice-agent turns and sub-500-millisecond real-time transcription. These vendor-reported latency claims have not been independently validated here.
The figures describe different parts of the system. After transcription, an application may still need to determine that the speaker has finished, run the reasoning model, execute a tool and begin producing speech. Fast transcription alone cannot establish the speed of a completed agent response.
Mistral says its end-of-turn prediction combines acoustic and semantic information. The system considers both how someone sounds and what they appear to be saying when deciding whether to respond.
The company’s real-time evaluation charts show a speed-accuracy tradeoff. On its reported 35-language FLEURS evaluation, the Mini real-time model records 7.4% error at the setting labeled 240 milliseconds, 6.4% at 480 milliseconds, 6.0% at 960 milliseconds and 5.7% at 2,400 milliseconds. Lower is better.
Those settings are points in Mistral’s transcription evaluation, not complete voice-agent turnaround times. A production test should separately measure recognition delay, end-of-turn behavior, time to first audible response and delays introduced by tools, using the intended network path and deployment region.
The Accuracy Charts Need Their Evaluation Conditions
Mistral’s published benchmark charts support a narrower conclusion than a blanket claim that Voxtral beats every competing speech model.
On the announcement’s 45-language FLEURS comparison, Small records a 4.0% inferred-language error rate, matching the displayed Microsoft AI Transcribe 2 result. Mini records 4.5%. The chart distinguishes results where the language is supplied from those where the model detects it automatically.
The reported aggregate uses word error rate for most languages, but character error rate for Japanese, Korean, Thai and Chinese. This macro-average across the selected language intersection is not an expected error rate for every call or language.
For diarization, Mistral reports results using concatenated minimum permutation word error rate, or cpWER. Small scores 18.4% and Mini 18.6% across the listed public datasets. That metric evaluates the transcript with speaker attribution, so it is not directly interchangeable with the plain transcription percentages.
Mistral presented these results; Zeniteq has not independently validated them. They justify testing both models for purchasing decisions, while leaving buyers to evaluate performance on their own accents, background noise, overlapping speakers and specialist vocabulary.
Private Deployment Is Offered, but Open Weights Are Unconfirmed
Mistral says the family is available through European, US and global API endpoints, with self-deployment options for on-premises infrastructure or private cloud environments. The announcement also emphasizes control over audio processing and retention.
Organizations evaluating where conversations can be processed still need answers beyond endpoint location. Buyers should establish retention settings, logging behavior, contractual commitments and the handling of any external tools called by an agent.
The self-deployment offer should not be described as an open-weight release. The supplied announcement does not establish publicly downloadable weights, a public license or unrestricted redistribution rights. Deployment availability and weight availability are separate questions.
Voxtral 3’s strongest practical contribution is the choice between buying components and buying an integrated conversation pipeline, with published rates for both. Builders with an existing voice architecture can evaluate transcription or TTS separately; teams starting a new service can assess the agent API. Measured task completion, language-specific performance and total operating cost under the intended deployment conditions should guide that choice. The fastest vendor latency figure is not enough.
