Voxtral 3: Transcription, Speech and Voice Agent Costs

Voxtral Speech 3 costs $13 per million characters for batch or realtime synthesis. For example, 100,000 billable characters cost $1.30. Transcription and Voice Agent rates are separate:

For transcription only, multiply billable API minutes by the matching rate.

Managed Voice Agent rates are quote-only. The launch does not define call-duration rounding, silence or hold-time billing, or whether telephony and tool charges are included. Confirm those terms before budgeting a complete call; do not add an Agent quote to the transcription estimate.

The audio overview lists five task-specific models as of October 25, 2026. Use the full model ID for the task you are integrating; the Voice Agent SDK combines services rather than adding a sixth speech-model ID. The launch page presents four model cards, so its card count is not the current API inventory.

The Mini Transcribe 3 and TTS 3 model pages carry Public Preview labels. API access and a self-deployment offer are different routes: contact Mistral for private-cloud or on-premises terms, and obtain the exact checkpoint and license before budgeting local operation. These pages do not establish downloadable open weights for the entire Voxtral 3 family.

For voxtral-mini-tts-3-0, the model page lists nine supported languages on October 25, while the launch article advertises 35+ languages and 50+ dialects. Confirm your intended language on the specific interface before choosing a voice or promising coverage. The TTS API guide also marks free-form style instructions unavailable at launch; its supported inline emotion and pace tags are a separate control.

Replace the old model ID using Mistral’s migration table, then test the request and response handling. Word-level timestamps are removed: request segment timestamps, since word returns a 400 error. Send all realtime settings before the first audio chunk, and let a diarization display retain overlapping segments.

For TTS, old emotion-specific preset voices remain accepted; inline emotion tags let one voice vary its delivery. Test saved or cloned voices and measure latency from your application again. The migration guide lists a rolling Small Transcribe alias and says the TTS latest alias changes within a month of GA; the explicit IDs above make the intended target clearer during a migration.

Use Mistral Studio to try the relevant task, then follow its documentation for API integration. For an agent, test interruptions, tool responses and a complete call before setting a call budget.