Mistral launches Voxtral 3 for AI voice agents

Mistral's Voxtral 3 can power AI voice agents for telephone and web conversations in Spanish and other languages, with options for keeping audio processing in Europe. Its complete voice agent API starts at $0.04 per minute, while individual transcription and speech-generation models are available separately.

Mistral introduced the technology on 20 October 2026, publishing a new family of speech models alongside a real-time voice agent SDK. The official announcement covers the technical capabilities, deployment options and prices.

For businesses, the decision involves more than selecting a natural-sounding voice. An operational agent needs access to the right systems, clear permissions and a reliable way to involve a member of staff when required.

A voice agent needs four separate capabilities

A working voice agent must listen, interpret, act and speak.

Listening starts with speech recognition. Voxtral Small Transcribe 3 supports more than 50 languages, including Spanish, and includes speaker diarisation to identify who is speaking.

Reasoning is handled by a language model that interprets the customer's request and decides what information is needed.

Taking action requires connections to business applications. Depending on the tools provided, the agent might check an order, retrieve account information or request an appointment.

Speaking uses text-to-speech generation. Voxtral Mini TTS 3 supports over 35 languages and 50 regional varieties.

Mistral has designed these components to work together in a streaming conversation rather than waiting for each stage to finish before the next begins.

Its Voxtral 3 announcement also describes interruption handling, allowing callers to speak while the agent is responding.

Consider an appointment-booking service. The voice agent can interpret a request and ask for missing information, but it still needs an authorised connection to the booking system to complete the task.

Speech recognition alone cannot provide that business functionality.

Does Voxtral 3 support Spanish from Spain?

Yes. Mistral supports Spanish speech recognition and synthesis, including a regional voice option for Spain.

This is useful for businesses serving Spanish-speaking customers who expect familiar pronunciation and regional conventions.

The transcription models also support context biasing, which allows developers to provide vocabulary that the system should recognise more reliably.

For example, product names or specialised terminology can be included in the recognition context.

Voxtral Mini Transcribe 3 is a 4-billion-parameter model intended for efficient batch transcription.

For live conversations, Voxtral Mini Transcribe Realtime 3 is designed to process speech with latency below 500 ms, according to Mistral.

The company also claims that its SDK can complete an agent conversation turn in under 1 second.

These figures describe the vendor's reported capabilities. They should not be treated as independently measured performance for a particular business deployment.

Network conditions, external tools and backend response times will all affect the actual experience.

How much does Voxtral 3 cost?

Mistral has published separate prices for its speech models and complete voice agent API.

Source: Mistral AI, 20 October 2026. All prices are in US dollars.

The distinction between individual model pricing and the complete agent API matters.

A transcription endpoint is not a complete customer-service agent. It converts audio into text but does not independently manage the rest of the conversation.

Similarly, a deployed voice agent may need telephony services, database connections, monitoring and ongoing maintenance.

These requirements should be costed separately where applicable, without assuming that individual model prices automatically describe the cost of the entire system.

Our article on the published costs of running an AI agent explains why model usage and system-level expenditure need separate consideration.

Can customer data remain in Europe?

Mistral provides European, US and global API endpoints, allowing organisations to select an appropriate hosted environment.

For greater deployment control, Mistral also offers arrangements for running the technology on company infrastructure or in a private cloud through its sales team.

These options are relevant to businesses concerned about the geographical location of customer conversations.

However, choosing an European AI provider does not automatically guarantee that every component of a deployed service processes data in Europe.

The full system may include a telephone provider, application backend, monitoring tools and external business integrations.

Each component needs to be included in the data-flow review.

Businesses should establish where recordings and transcripts are processed, whether they are retained and who can access them.

According to Mistral, self-deployment allows organisations to control audio handling and retention within their own environment.

This is a technical option that needs to be matched with the organisation's operational and data-protection requirements.

LiveKit, SIP and WebRTC make integration possible

The Voxtral 3 voice agent SDK is built on LiveKit and supports established communication technologies.

SIP enables integration with compatible telephone systems and call infrastructure.

WebRTC enables live audio conversations in browsers and applications.

This allows developers to build telephone agents and website voice assistants using the same general speech pipeline.

The SDK also separates transcription, reasoning and speech synthesis into individual components.

A developer can replace a compatible language model or voice component without necessarily rebuilding the entire application.

TestingCatalog's report from 21 October 2026 confirms the SDK architecture and communication options.

Modularity is useful for long-term maintenance, although changing a component still requires testing to make sure the resulting conversations behave as intended.

What should a business check before deployment?

Human handover deserves particular attention in customer service.

An agent should not leave a caller trapped in an automated conversation when the request cannot be resolved.

For companies operating in Spain, the relevant requirements are discussed in our article on AI customer-service agents and Law 10/2025.

What should you decide this week?

Choose one specific customer interaction, such as checking an order or booking an appointment, and document the systems it needs to access. Confirm the required processing region, review the published usage prices and test the conversation in Spanish before committing to a wider deployment. If you are considering a telephone or website voice agent, tell us which customer enquiries you want to handle so we can assess the integration requirements.