Voxtral 3 separates interview transcripts from live voice agents

An hour-long interview needs a transcript. A narrated explainer needs a voice. A spoken assistant needs to listen and respond while someone is talking. Mistral’s October 20 Voxtral 3 launch covers all three, with separate models and prices rather than one interchangeable audio endpoint. Release announcement.

For a creator editing a recorded conversation, start with the file-transcription route. The Mini Transcribe 3 card lists $0.0035 per minute and labels the model Public Preview. Mistral’s launch table lists Small transcription at $0.004 per minute. An illustrative 60-minute file therefore has a transcription subtotal of $0.21 with Mini or $0.24 with Small, before other processing. Those are calculations, not invoices or quality measurements.

The small price gap makes the correction work matter

Saving three cents on that hour is less useful if the transcript requires much more cleanup. Before moving a podcast workflow, propose a short comparison using the same difficult excerpt: two speakers, an interruption, a product name and a spoken date. Keep the reference text and count the corrections that affect meaning. We have not run this comparison.

The audio overview separates recorded-file models from live transcription and speech generation. A stream of captions is a different integration from uploading a finished interview. Choose the workflow first, then check the specific model’s supported vocabulary, speaker labels and limits.

Narration has its own controls and bill

The launch lists speech generation at $13 per million characters. Its full voice-agent API has separate per-minute rates; the file-transcription subtotal does not buy a complete spoken assistant. Mistral’s speed and accuracy comparisons remain vendor reports, not tests by CreatorStack. Launch pricing and SDK.

For a voiceover, the current TTS table documents emotion and pace tags plus phoneme-based pronunciation. It also says whole-request natural-language style instructions are not available at launch. Use those documented controls rather than planning an edit around the broader promotional description.

Start with a paragraph containing your hardest name and the intended delivery. Listen for meaning and pronunciation before generating the full script. Our AI output evaluation guide helps define that acceptance check; the podcast repurposing workflow shows where the transcript fits in an edit.