Latest in AI
Mistral Voxtral 3: API Pricing, Voice Agents and Migration Changes
Mistral combines new speech models with a voice-agent SDK, but developers should check the launch-day pricing and integration details before switching.
Mistral launched Voxtral 3 on October 20, 2026, combining updated transcription and speech-generation models with a voice-agent SDK. The release targets applications that need to listen, reason, call tools and speak during a live conversation. It also changes integration details that existing Voxtral developers should check before migrating.
The practical shift is from choosing individual audio models to assembling a complete conversational workflow. That is useful, but a voice agent's reliability still depends on what happens between hearing a request and acting on it.
Voxtral 3 API pricing
Mistral's launch announcement lists these prices:
| Component | Announced API price |
|---|---|
| Voxtral Mini Transcribe 3, batch | $0.0035 per minute |
| Voxtral Small Transcribe 3 | $0.004 per minute |
| Voxtral Mini Transcribe Realtime 3 | $0.008 per minute |
| Voxtral Speech 3 | $13 per million characters |
There is a launch-day discrepancy worth checking: the linked Voxtral TTS 3 model card lists $16 per million characters, rather than $13. It also lists nine languages, while the announcement describes more than 35 languages and 50 dialects. These official pages do not yet agree, so neither the final speech price nor the supported language set should be assumed from the headline alone.
The TTS card labels the model Public Preview and exposes the model ID voxtral-mini-tts-3-0. Confirm the applicable rate and required language in your deployment before estimating production costs.
For transcription alone, 1,000 minutes works out to $3.50 for Mini batch, $4 for Small or $8 for realtime at the announced rates. Those calculations exclude generated speech, the reasoning model, tools and any telephony or transport charges. Transcription pricing is not the total cost of a voice agent.
How the voice-agent SDK works
The voice-agent documentation describes a pipeline with separate transcription, reasoning and speech stages. The SDK builds on LiveKit and provides Python and TypeScript bindings. Users can interrupt while the agent is speaking, and the reasoning model can call tools during the conversation.
The separation is useful for application design. You can choose a model for each stage rather than treating the entire interaction as one opaque speech model. Saved agents are versioned configurations, while sessions represent individual conversations; pinning an agent version helps keep production behaviour stable during testing.
Phone support has an important restriction. SIP telephony is available to enterprise customers through their account team at launch. Client-side tools do not run on phone calls, so telephone agents need server-side tools for their actions.
For a service desk, this is a meaningful implementation difference. A function that works inside a browser demo is not automatically available when the same agent answers a telephone call. Tool placement needs to be part of the deployment plan.
What changes from Voxtral 2
The migration guide identifies changes beyond replacing a model name.
Word timestamps are no longer supported. Requests must use segment timestamps; asking for word-level output returns an error. That matters for applications with word-highlighted playback or tightly aligned captions.
Realtime settings also lock when the first audio chunk arrives. Send the complete session configuration before streaming audio, rather than updating it midway through a call. Interfaces must handle overlapping diarized segments instead of assuming speakers always take perfectly separate turns.
Context biasing now accepts multi-word phrases, with free-form context supported offline. This can help with product names and specialised vocabulary, but the supplied context still consumes the model's available context window.
On speech generation, emotion controls move toward inline tags. The migration page explicitly marks style instructions as unavailable at launch, so a feature described broadly in the announcement should not be treated as a working integration until the relevant documentation supports it.
A realtime limitation to test early
The speech-to-text guide says realtime diarization and turn detection cannot be enabled in the same session. Configuring both returns an error. Each feature is supported, but not together in one session.
That is particularly relevant for agents expected to identify speakers while deciding when to respond. Choose the conversation behaviour you need, then test the supported configuration with realistic interruptions, background speech and domain-specific terms.
Voxtral 3 makes more of the voice stack available as one developer workflow. The useful measure is whether that workflow completes the right action, recovers from a misheard detail and respects a caller's interruption. A natural voice helps, but those behaviours determine whether people trust the service.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn launches and complex product updates into clear stories that developers and founders want to read and share.