Voxtral 3 makes the reasoning model a pricing decision

What shipped

Mistral’s Voxtral 3 announcement covers three connected products: speech-to-text models, text-to-speech models, and a LiveKit-based SDK for building real-time voice agents. The SDK brings transcription, reasoning, tool calls, and speech generation into one streaming pipeline.

The distinction between the models and the agent API matters. Mistral lists separate prices for batch transcription, realtime transcription, speech generation, and the voice-agent API. Engineers can evaluate the speech components individually or start with the agent offering. The announcement is not limited to a transcription model or a speech-generation endpoint.

Mistral says Voxtral 3 is available through its European, US, and global API endpoints. It also offers self-deployment, with on-premises and private-cloud deployments requiring contact with its team. Those are separate entry points: hosted API use is available through the listed endpoints, while private deployment starts with a conversation with Mistral.

The company also provides Studio playgrounds for voice agents, transcription, speech generation, and voice cloning. That gives engineers a starting point for testing each part of the release without treating the full voice-agent pipeline as the only product.

The headline price needs its qualification attached. Mistral charges $0.04 per minute for the voice-agent API when Mistral Small 4 handles reasoning. With another reasoning model, the listed rate is $0.08 per minute. Model selection therefore affects the agent rate directly, not just the reasoning behavior.

How the streaming loop works

Mistral describes the SDK as a LiveKit-based pipeline with three stages:

• Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished speaking.
• A general model, such as Mistral Small, reasons and calls tools.
• Voxtral Speech 3 generates a reply using the tone of the conversation so far.

According to Mistral, those stages stream rather than waiting for the preceding stage to finish. The company claims each turn completes in under one second. It also says the agent stops talking when the caller interrupts it.

Those are claims about the assembled voice experience, not just transcription speed. Mistral separately reports sub-500ms latency for realtime transcription. Keep the two figures separate when evaluating the release: one describes the transcription model, while the other describes the agent’s turn. The supplied material does not establish how either figure changes with a particular reasoning model or tool call.

For integration, Mistral says agents can use function calling, web search, and retrieval over an organization’s systems. Its examples include verifying a caller’s identity, checking a bill, and booking an appointment. These are examples of tool-backed agent behavior, rather than evidence of a particular workflow’s accuracy or completion rate.

The transport options cover phone and application experiences. Mistral says agents can answer and place calls over SIP, run in browsers and apps over WebRTC, and connect to existing telephony, IVR, and contact-center software.

What the transcription claims establish

Mistral positions its transcription models as state of the art. It says Voxtral Small Transcribe 3 outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe, with strong results across diarization and multilingual transcription, including English, French, and Spanish.

That comparison remains a vendor claim. The supplied text includes benchmark descriptions but no numerical scores. It therefore supports reporting which competitors Mistral names, but not quantifying the lead or deciding how large an improvement an application should expect. There is no supplied numerical basis for ranking those models on your workload.

The Mini transcription variant has 4 billion parameters, according to Mistral. The company lists both Mini and Small batch transcription offerings, along with realtime Mini transcription. The realtime model is positioned for agent use and is reported to deliver sub-500ms transcription latency with diarization and context biasing.

Language coverage needs to stay attached to the relevant model description. Mistral describes its transcription models as supporting more than 50 languages. Its realtime-model description reports more than 35 languages, while its realtime product listing states 40-plus. Do not turn those different statements into one universal language count.

Mistral also describes context biasing through key terms and free-form text. It says this can improve recall of technical vocabulary without fine-tuning. That is a concrete feature to evaluate with domain-specific terms, alongside diarization and end-of-turn detection. Keep that test separate from the broader benchmark claim: supplying vocabulary is a capability Mistral describes, while comparative accuracy still needs evidence relevant to the intended workload.

Speech generation and voice adaptation

Mistral offers Voxtral Speech 3 for batch and realtime text-to-speech. It reports support for more than 35 languages and more than 50 dialects, at $13 per million characters.

The company claims state-of-the-art expressiveness and instruction following. Mistral lists contextual emotion and delivery, custom instruction following, pronunciation support, and voice cloning or adaptation from at least 30 minutes of audio. The announcement also names a vocabulary library.

These features provide several separate things to test, rather than accepting the announcement’s broad expressiveness claim as a single quality measure.

The cloning feature also has a stated input requirement. Mistral describes 30-plus minutes of sampled audio, so that is the supplied starting point for evaluating voice adaptation. The announcement does not provide a numerical quality result for a cloned voice in the supplied text.

Mistral says Voxtral Speech 3 costs less than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. The supplied material does not include those competitors’ prices. The defensible comparison is therefore limited: Mistral lists its own $13-per-million-character rate and claims a cost advantage. It does not provide enough comparative pricing here to quantify that advantage.

The rate card has a reasoning-model dependency

The supplied pricing separates agent minutes, transcription minutes, and generated characters:

• Voice-agent API with Mistral Small 4 reasoning: $0.04 per minute.
• Voice-agent API with another reasoning model: $0.08 per minute.
• Mini batch transcription, 4B parameters: $0.0035 per minute.
• Small batch transcription: $0.0040 per minute.
• Realtime Mini transcription: $0.008 per minute.
• Voxtral Speech 3: $13 per million characters.

The agent price is explicitly conditional. An evaluation that uses another reasoning model should carry the $0.08-per-minute rate, rather than the $0.04 figure in the launch headline. Keep the reasoning-model choice visible in any internal comparison of the agent offering.

The standalone component rates answer different questions. Batch transcription has its own Mini and Small prices. Realtime Mini has a separate per-minute price. Speech generation uses characters rather than minutes. Preserve those units when comparing options, rather than presenting the release as having one general voice price.

The supplied material does not provide a complete billing breakdown connecting every component rate to the agent rate. Do not infer additional charges, or assume a particular inclusion, from the listed rates alone. Use the published agent rates for the stated agent configurations and the standalone rates for their named services, then confirm billing details for the intended integration.

Private deployment is another item to resolve separately. Mistral offers on-premises and private-cloud deployment through its team, but no private-deployment price is supplied here. The hosted rate card is not a quoted price for that deployment path.

What to evaluate before integrating

The release is relevant to engineers building phone agents, browser or app voice experiences, transcription systems, and speech-generation features. Mistral offers the speech models alongside its SDK for real-time voice agents.

For a voice agent, start with the streaming behavior Mistral emphasizes. Test turn completion, caller interruption, and the required function calls. Use SIP or WebRTC according to the target experience. Keep the selected reasoning model in the evaluation record because it determines which published agent rate applies.

For transcription, focus on the intended languages, diarization, technical vocabulary, and whether the workload requires batch or realtime processing. Mistral’s named competitor comparisons are useful context, but the supplied benchmark text does not quantify a performance gap.

For speech generation, Mistral lists instruction following, pronunciation support, contextual delivery, and support for more than 35 languages and more than 50 dialects. If voice adaptation is required, start from Mistral’s stated 30-plus-minute audio requirement rather than assuming a shorter sample is supported.

Finally, distinguish the deployment decision from the model decision. Mistral lists European, US, and global hosted endpoints, while on-premises and private-cloud deployments require contacting its team. Studio provides the initial playgrounds for the four advertised experiences.

Try it

Start in Mistral Studio’s voice-agent, transcription, speech-generation, or voice-cloning playgrounds, or use the Mistral API. For on-premises or private-cloud deployment, contact Mistral. The supplied sources contain no executable commands.

Sources

• https://mistral.ai/news/voxtral-3/