Mistral Voxtral 3 Review: Voice Agent SDK Features, Pricing and Limits

Mistral Voxtral 3, released on October 20 2026, combines speech-to-text, text-to-speech and a voice agent SDK built on LiveKit. The system handles end-to-end spoken conversations with tool calling in a single streaming pipeline. Organizations evaluating real-time agents can now compare its reported latency, language coverage and per-minute costs against separate component stacks.

Mistral Voxtral 3 SDK Features and Language Support

The SDK integrates transcription, reasoning via any LLM such as Mistral Small 4, and TTS output. It supports barge-in, end-of-turn detection that combines acoustic and semantic signals, and diarization for multi-speaker calls. SIP telephony and WebRTC connections are built in, allowing direct connection to existing contact-center infrastructure without additional gateways.

Transcription covers 50 languages, with realtime streaming available in more than 35. TTS operates in 35 languages and over 50 dialects. Context biasing accepts up to 30 minutes of reference audio for voice cloning, while instruction tags control emotion and pace during synthesis. Tool calling permits multi-step workflows such as identity verification, bill lookup and appointment booking through function calls or retrieval.

Reported performance includes macro WER between 4.0 and 4.1 percent on FLEURS 45 for the Small Transcribe model and sub-500 ms latency in typical streaming configurations. The architecture allows swapping the underlying LLM, voice model or language without rebuilding the pipeline, which reduces integration time for teams already using Mistral models.

Voxtral 3 Pricing, Tool Calling Limits and Deployment Options

Pricing is usage-based and billed per minute or per character. Batch transcription costs $0.0035 per minute for the Mini model and $0.0040 per minute for Small. Realtime transcription runs at $0.008 per minute. TTS is charged at $13 per million characters. Voice agent sessions that include Mistral Small 4 reasoning cost $0.04 per minute; sessions using other reasoning models cost $0.08 per minute. Regional surcharges of 10 percent may apply outside primary EU and US endpoints.

Exact rate limits on concurrent tool calls and session quotas are not published in the initial announcement. Enterprise customers must contact Mistral directly for production quotas and self-hosting terms. The service runs on Mistral API endpoints in the EU and US, with self-deployment available for organizations that require data residency controls. No open weights have been released for the Voxtral 3 family, unlike earlier versions.

Compared with separate ASR plus TTS stacks, the integrated pipeline removes the need to manage synchronization between components. Teams that already run Mistral Small 4 for text agents can reuse the same model inside voice sessions, keeping context length and tool definitions consistent. Self-hosted options allow retention policies that satisfy strict EU data rules, though exact hardware requirements and support SLAs remain available only on request.