Mistral Voxtral 3: Voice AI Models, Pricing and How to Get Started

Mistral AI introduced Voxtral 3 on October 20, 2026, expanding its speech technology into a system for building conversational AI agents. The release includes new transcription and text-to-speech models plus a software development kit that connects listening, reasoning, tool use and spoken replies. Mistral says its transcription lineup covers more than 50 languages, while its voice-agent offering starts at $0.04 per minute when paired with Mistral Small 4.

The announcement is more than another speech-to-text model. Voxtral 3 aims to address a practical problem: making a voice assistant that can understand a caller, handle interruptions, retrieve information and complete an actual task without forcing a developer to assemble every component independently.

Here is what the new models do, how the advertised pricing works and which deployment claims deserve a closer look.

What is Voxtral 3?

Voxtral 3 is a family of speech models and an agent-building toolkit, not one general-purpose chatbot. Its components serve different jobs: transcribing recorded audio, understanding speech as it happens, generating spoken responses and coordinating a conversation with a reasoning model.

In its official launch announcement, Mistral describes applications ranging from customer service and telephone bookings to hands-free assistants. A voice agent can listen to a request, consult an external system through a tool call and speak the result back to the user.

That does not mean every business action works automatically. Booking an appointment, for example, still requires a compatible booking system, authentication, appropriate permissions and safeguards against unintended changes.

Voxtral 3 models: which one does what?

Mistral's release separates the underlying speech workloads. Choosing a model depends on whether an application processes recordings after the fact or must respond during a live conversation.

Batch and higher-accuracy transcription

Voxtral Mini Transcribe 3 is positioned as the inexpensive batch option. It processes existing audio and supports multilingual speech recognition, speaker diarization—identifying who spoke—and contextual vocabulary to help recognize specialized terminology. Mistral lists $0.0035 per minute at launch.

Voxtral Small Transcribe 3 targets more demanding transcription and related audio-understanding work, at a listed $0.004 per minute. Mistral claims strong benchmark results, but organizations should test it against their own accents, microphone quality, jargon and real recordings before accepting performance comparisons at face value.

Streaming transcription for live agents

Voxtral Mini Transcribe Realtime 3 is built for conversations in progress. It supports streaming transcription and end-of-turn detection, which helps an application estimate when a person has finished speaking. Its published rate is $0.008 per minute.

Mistral's technical model documentation describes configurable streaming latency and WebSocket integration. That makes this variant more relevant to a telephone assistant or live voice interface than a nightly transcription job.

Speech synthesis and voice adaptation

Voxtral Speech 3 generates audio responses and supports expressive delivery, with controls intended to improve pacing, pronunciation and emotional tone. Mistral's announcement lists $13 per million characters for the speech service.

The launch also discusses voice adaptation from sample recordings. Organizations considering cloning or adapting a voice should secure authorization from the speaker and test how the system handles impersonation risks, consent and retention policies.

Model documentation and commercial pricing pages can use different product names or updated rates; developers should check the exact model identifier and live billing terms before deployment.

How does the Voxtral 3 voice-agent SDK work?

The new SDK uses LiveKit to coordinate the components of a live conversation. Mistral explains that realtime transcription detects speech, a separate language model such as Mistral Small can interpret requests and call tools, and the speech model generates the answer.

The design supports interruptions: if a caller starts talking while the agent is responding, the system can stop speaking and process the new input. This matters more to a natural telephone conversation than simply producing audio that sounds convincing.

The SDK is also intended to connect with real systems through functions, retrieval and search. According to Mistral, integrations can use SIP for telephone calls and WebRTC for browser or application audio.

In practice, teams must still design what the agent is allowed to do. A customer-support bot may read account details, but issuing refunds or changing an appointment should require explicit authorization, auditing and a safe failure path.

Voxtral 3 API pricing: how much does it cost?

Mistral's October 20 launch page lists the following headline rates:

These are vendor-advertised launch prices, not an independent estimate of a deployed contact center's total costs. Calls to external tools, telephony service, storage and other infrastructure may affect the bill.

For a simple illustration, 1,000 billable voice-agent minutes at $0.04 per minute would equal $40 at the listed API rate before any other applicable charges. A service charging $0.08 per minute for that usage would total $80. Actual invoices depend on how usage is measured and which components a deployment selects.

Can Voxtral 3 run locally or privately?

Mistral says Voxtral 3 is available through its hosted API, including European endpoints, and that organizations can arrange self-deployment or private-cloud operation. That could interest teams processing sensitive recordings or subject to data-residency requirements.

But private deployment is not the same as a confirmed free download for a home GPU. Mistral directs organizations interested in running the full stack on their own infrastructure to contact its team. Check the availability, licensing, hardware requirements and support terms for the specific model rather than assuming every Voxtral 3 component is open-weight.

Keeping audio within a chosen environment can help control data handling, but it does not by itself resolve privacy compliance. Retention, recording consent, access control and incident response remain the operator's responsibility.

Should developers try Voxtral 3 now?

For teams already working on voice agents, Voxtral 3 gives a new combination to evaluate: inexpensive batch transcription, a realtime speech option, expressive output and an SDK for connecting conversations to tools.

A useful pilot should measure:

• Recognition quality with real speakers, overlapping dialogue and domain-specific names.
• End-to-end response latency, not just one model's advertised benchmark.
• Interruptions and recovery, particularly when an agent mishears a number or address.
• Tool reliability and approval controls, including how mistaken actions are prevented.
• Total operating cost using the actual telephony, models, infrastructure and usage profile.

The key announcement is clear: Mistral is now competing to provide the complete voice-agent workflow, rather than only a transcription endpoint. What remains to be demonstrated independently is how reliably that experience performs across real languages, noisy environments and complex customer-service tasks.

Developers can review the Voxtral 3 announcement and the realtime transcription documentation before choosing the components for a test.