# Mistral Voxtral 3 targets OpenAI and ElevenLabs with EU-sovereign voice agents

> The French lab pairs state-of-the-art transcription with a real-time voice agent SDK, aiming to make EU data residency a selling point against US rivals.

*Published 2026-10-26 · By Marcus Vance*

Mistral Voxtral 3 is a family of speech-to-text and text-to-speech models plus a voice agent SDK released Oct. 20, 2026, that lets developers build real-time multilingual voice agents through Mistral's EU, US, or global API endpoints.

## What is Voxtral 3 and why does it matter?

Mistral released Voxtral 3 on Oct. 20, 2026, a coordinated launch of transcription models, a text-to-speech model, and a voice agent SDK that puts the French AI company in direct competition with OpenAI and ElevenLabs. The release targets developers building voice agents for customer service, call centers, and real-time enterprise communications, according to Mistral's announcement.

The product family covers two distinct workloads. Batch transcription converts recorded calls and media files into text at a lower price point, while real-time transcription supports live conversations with streaming output. Voxtral Mini TTS 3 generates speech, and the voice agent SDK orchestrates the full loop from audio input to spoken response.

The launch arrives as voice agents move from experimental demos to production deployments in contact centers. Enterprises that once used interactive voice response trees are testing systems that can understand callers, query internal systems, and respond naturally. Mistral is betting that those deployments in Europe will prefer a vendor that can keep audio data inside EU jurisdiction.

## What does Voxtral 3 include?

The Voxtral 3 family has five components, each with a distinct role in the voice pipeline. Mistral describes the lineup as a single toolkit rather than a collection of disconnected models, with the SDK acting as the integration layer.

- Voxtral Small Transcribe 3, Mistral's highest-accuracy batch transcription model.
- Voxtral Mini Transcribe 3, a 4B-parameter batch transcription model optimized for cost.
- Voxtral Mini Transcribe Realtime 3, a streaming model for live conversation transcription.
- Voxtral Mini TTS 3, a text-to-speech model priced at $13 per million characters.
- A voice agent SDK built on LiveKit that combines listening, reasoning, tool calling, and speaking in one pipeline.

Parameter count affects deployment economics. The Mini Transcribe 3 uses 4B parameters, according to Mistral, which makes it lighter than the Small variant and better suited to high-volume workloads where per-minute price is the binding constraint.

The separation between batch and real-time pricing reflects different infrastructure costs. Streaming transcription must return tokens continuously as audio arrives, which requires more compute per minute than processing a pre-recorded file. Mistral's pricing page lists real-time transcription at more than double the batch rate for the Mini model.

## How does the Voxtral 3 voice agent SDK work?

The SDK, built on LiveKit, combines listening, reasoning, tool calling, and speaking into a single real-time streaming pipeline, according to Mistral. LiveKit provides the WebRTC transport layer, managing audio streams, session state, and network connectivity, while Mistral supplies the speech models and orchestration logic.

From a developer's perspective, this collapses a multi-vendor stack into one integration. A typical voice agent today requires an automatic speech recognition provider, a large language model for reasoning, a separate tool-calling layer, and a text-to-speech provider. Voxtral 3 packages those steps into one pipeline, with Mistral Small 4 as the reasoning model for voice agent sessions, according to Mistral's pricing documentation.

Tool calling is the feature that distinguishes voice agents from simple speech assistants. The SDK lets a model invoke external APIs mid-conversation, such as looking up an order status or booking an appointment, and then speak the result. Mistral's announcement describes this as bringing listening, thinking, tool calling, and speaking into a single stream.

## How much does Voxtral 3 cost?

Mistral's API pricing page lists the audio models with per-minute and per-character rates. Batch transcription starts at $0.0035 per minute for Voxtral Mini Transcribe 3 and rises to $0.004 per minute for Voxtral Small Transcribe 3. Real-time transcription is priced at $0.008 per minute for Voxtral Mini Transcribe Realtime 3.

*Voxtral 3 pricing as listed by Mistral AI*

| Model | Listed price | Use case |
| --- | --- | --- |
| Voxtral Mini Transcribe 3 | $0.0035 per minute | Batch transcription |
| Voxtral Small Transcribe 3 | $0.004 per minute | High-accuracy batch transcription |
| Voxtral Mini Transcribe Realtime 3 | $0.008 per minute | Streaming transcription |
| Voxtral Mini TTS 3 | $13 per million characters | Text-to-speech |
| Voice agent with Mistral Small 4 | $0.04 per minute | Full voice agent runtime |

Those numbers put Voxtral 3 at the low end of commercial speech pricing, though Mistral's announcement does not include a direct comparison with OpenAI or ElevenLabs rate cards. The $0.04 per minute rate for a voice agent using Mistral Small 4 bundles the reasoning component into the runtime cost, which simplifies budgeting for enterprises that want a predictable per-call price.

Per-character pricing for text-to-speech is standard in the industry. Voxtral Mini TTS 3's $13 per million characters means a 1,000-character response costs about 1.3 cents, making synthetic speech inexpensive relative to the per-minute cost of running the full voice agent loop.

## Which languages does Voxtral 3 support?

The transcription models support 50 languages, including all 24 official EU languages and many low-resource Asian and Indic languages, according to Mistral. Language coverage is a core differentiator for the release because much of the voice AI market has concentrated on English and other high-resource languages.

The EU language angle is deliberate. All 24 official EU languages are included, from Maltese to Finnish, which is rare among commercial speech APIs. For public-sector buyers in Europe, the ability to handle every member state's language in one model reduces the need to assemble a patchwork of regional vendors.

## How can enterprises deploy Voxtral 3?

Mistral offers Voxtral 3 through API endpoints in the European Union, the United States, and a global region. The EU endpoint is the default choice for organizations subject to GDPR and sector-specific data rules, while the US endpoint serves customers with existing US infrastructure. Self-deployment is available for enterprises that want to run the models inside their own environment.

The absence of public weights distinguishes Voxtral 3 from Mistral's earlier open-weight releases. Mistral has not published the model on Hugging Face, and there is no indication of a community edition. Instead, the company is treating speech as a commercial product line, with the SDK and API as the primary distribution channels.

That decision has trade-offs. Open weights would accelerate community experimentation and third-party integrations, but they would also make it harder for Mistral to control quality, enforce safety guidelines, and capture revenue. The enterprise focus of Voxtral 3 suggests Mistral is prioritizing commercial contracts over ecosystem adoption.

## What do Mistral's benchmarks show?

Mistral's benchmarks position Voxtral Small Transcribe 3 as state-of-the-art in multilingual transcription, diarization, and context-biasing. The company says it outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe on those tasks. Diarization, the ability to distinguish speakers in a conversation, is critical for call-center analytics and legal use cases.

Benchmark claims from model vendors are self-reported unless a third party validates them. Mistral does not say in the announcement whether an independent evaluator ran the comparisons. Buyers will likely run their own evaluations on representative call audio before committing, particularly for diarization accuracy.

Context-biasing, another area where Mistral claims state-of-the-art results, lets a model favor domain-specific terms such as product names, ticket numbers, and proper nouns. That matters in enterprise telephony, where generic transcription models frequently mangle company-specific vocabulary.

## How is Mistral positioning the release?

Mistral frames Voxtral 3 as a response to fragmentation in voice AI tooling. The company's announcement argues that developers need a single pipeline rather than separate components for speech recognition, reasoning, and speech synthesis.

> To address this, we introduce Voxtral 3: a new family of state-of-the-art speech-to-text and text-to-speech models, alongside a voice agent SDK that brings listening, thinking, tool calling, and speaking into a single real-time streaming pipeline.Mistral AI

The announcement pairs the technical release with a commercial argument: enterprises can keep voice data inside EU infrastructure while using a model family designed for multilingual European use. Mistral's API endpoints in the EU, US, and global regions give customers a choice about where audio and transcripts are processed.

## What are the implications for the European voice agent market?

The EU endpoint is the clearest differentiator against US-based rivals. Banks, insurers, hospitals, and public agencies across Europe face data-protection rules that make it difficult to send customer conversations to US-hosted models. Voxtral 3's availability through EU endpoints lets those organizations use a speech system without a transatlantic data transfer.

Self-deployment adds another layer of control. Enterprises that require full data locality can run Voxtral 3 on their own infrastructure under a commercial license, even though Mistral has not released public weights. This hybrid model keeps the technology proprietary while addressing the sovereignty concerns that dominate European procurement conversations.

Price competition is the second lever. At $0.0035 per minute for batch transcription, Voxtral 3 makes it economical to transcribe entire call histories rather than sampled interactions. Large contact centers process millions of minutes per month, so a fraction of a cent per minute translates into meaningful annual savings.

The broader effect is downward pressure on speech AI costs. Mistral's entry into the voice agent market with an EU-native option gives procurement teams a credible alternative to US platforms and a benchmark for negotiating rates.

## What are the risks for Voxtral 3?

Mistral faces execution risks. Self-reported benchmark leadership must survive independent testing, and the absence of public weights limits third-party scrutiny. The voice agent market is also crowded, with incumbents that have larger distribution and established developer ecosystems. Mistral's advantages are geographic and linguistic rather than scale.

The company's choice of Mistral Small 4 as the reasoning engine for voice agents creates a natural upsell path. Developers who start with Voxtral transcription can add the voice agent SDK, which pulls in Mistral's LLM for tool calling and dialogue management. Each step increases usage of Mistral's broader model catalog, but it also means customers must trust a single vendor across the entire speech stack.

## What comes next for Voxtral 3?

Mistral has not published a specific roadmap for the Voxtral family beyond the initial release. The architecture separates speech models from the SDK, so Mistral can update individual components without forcing developers to rework integrations. That flexibility is useful in a market where latency, voice quality, and language coverage improve quickly.

Competitive pressure will likely shape the roadmap. OpenAI has been expanding its real-time speech API, and ElevenLabs has built agentic voice products around its text-to-speech models. Mistral's answer is to compete on price, language coverage, and EU data residency rather than on the sheer size of its models.

The immediate test for Voxtral 3 is adoption in European call centers and public-sector contact points. If the combination of 50-language coverage, low per-minute pricing, and EU endpoints wins enterprise contracts, Voxtral 3 will establish Mistral as a default sovereign voice platform. If not, the release will remain a technical showcase.

Mistral's broader strategy ties Voxtral 3 to its existing LLM lineup, with Mistral Small 4 serving as the reasoning engine for voice agent sessions. That coupling means developers who build on the LiveKit SDK and Mistral's speech models are adopting a full stack, not a single component.

## Sources

1. [Voxtral 3 introduces a voice agent SDK and state-of-the-art speech models; the Mini variant has 4B parameters; Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription, diarization and context-biasing; Voxtral Mini TTS 3 is priced at $13 per million characters.](https://mistral.ai/news/voxtral-3/)
2. [Audio models section lists Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Small Transcribe 3 at $0.004 per minute, Voxtral Mini Transcribe Realtime 3 at $0.008 per minute, Voxtral TTS 3 at $13 per million characters, and voice agents using Mistral Small 4 at $0.04 per minute.](https://mistral.ai/pricing/api/)

---
Source: https://aiintelreport.com/frontier-models/mistral-voxtral-3-eu-sovereign-voice-agents
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt