Mistral Voxtral 3 Pricing 2026: Transcription Costs, 50 Languages and Voice Agent SDK Limits

Mistral Voxtral 3 pricing starts at $0.0035 per audio minute for batch transcription. The family includes separate STT, TTS, and LiveKit-based voice agent components released on October 20, 2026.

Business users evaluating real-time voice pipelines need clear data on per-minute costs, language coverage, and SDK constraints before committing to production workloads.

Mistral Voxtral 3 Pricing Breakdown

Batch transcription runs through two models. Voxtral Mini Transcribe 3 costs $0.0035 per audio minute. Voxtral Small Transcribe 3 costs $0.004 per minute and adds speaker diarization plus context biasing.

Streaming transcription uses Voxtral Mini Transcribe Realtime 3 at $0.008 per minute with sub-500 ms latency and end-of-turn detection. The full voice agent pipeline, which includes Mistral Small 4 reasoning, lists at $0.04 per minute or $0.08 per minute when paired with another reasoning model.

TTS pricing sits at $13 per million characters for both batch and realtime output. These rates apply to Mistral’s EU, US, and global API endpoints. Self-hosted deployment requires direct sales contact; no public weights are available.

Compared with earlier Mistral audio releases, the 2026 tier adds per-minute billing for realtime agents and removes open-weight availability. Developers running high-volume contact-center traffic should budget for the higher realtime and agent rates rather than batch-only estimates.

Language Support and Transcription Features

Batch models handle 50 languages, covering all 24 official EU languages plus several low-resource Asian and Indic languages. Realtime transcription supports 45 languages and dialects. TTS covers 35 languages with 50 dialects; 18 languages include preset voices with native accents while others rely on voice cloning.

STT features include speaker diarization limited to eight reference clips, context biasing up to a 65k-token window, and end-of-turn detection. TTS adds emotion, pace, and emphasis parameters. Pro Voice cloning requires roughly 30 minutes of source audio.

Independent tests on Gulf Arabic, English, and Urdu voice notes produced competitive accuracy but did not lead public benchmarks. Mistral reports Small Transcribe 3 at 4.0–4.1 % WER on FLEURS 45 and 2.1 % AA-WER on Artificial Analysis non-streaming tests. Realtime variants show higher word error rates in some conditions.

Voice Agent SDK Limits and Practical Constraints

The LiveKit-based SDK streams audio through listening, reasoning, tool calls, and speaking in a single pipeline. Claimed turn latency stays under one second. SIP telephony integration remains enterprise-only at launch and focuses on inbound calls.

Public preview status means validation errors on invalid inputs and possible changes to rate limits. No separate tool-calling quotas appear beyond the underlying LLM limits. Context biasing shares the 65k-token window with the main transcription context.

Teams needing sovereign EU hosting can route calls through Mistral’s regional endpoints. For lower-latency or multilingual contact-center workloads, compare against ElevenLabs Eleven Realtime and existing Whisper-based pipelines before scaling.