Voxtral 3 FinOps: price audio by the minute and character

Updated October 27, 2026 · first published October 27, 2026

Voxtral 3 is priced on audio minutes for transcription and generated characters for speech. Mistral's changelog dates the public-preview release to October 20, 2026; its standalone announcement is dated October 21. The published rates make it possible to forecast audio processing directly, but a voice agent can add a separate per-minute charge for the managed pipeline and its reasoning model.

Four launch models, three billing meters

Mistral listed four public-preview models at launch: offline Mini and Small transcription, realtime Mini transcription, and Mini TTS. Its current pricing table gives these API rates in USD:

These are different meters. Do not convert audio minutes to text tokens or TTS characters using a single blended rate. For speech generation, the bill follows input characters; for transcription, it follows audio duration.

Budget examples you can audit

At the published rate, 10,000 recorded minutes cost $35 with Mini Transcribe or $40 with Small. The $5 difference buys the higher-accuracy tier, so compare cost per accepted transcript, not only cost per processed minute. Ten thousand realtime minutes cost $80.

TTS at $13 per million characters costs $130 per 10 million characters. A six-minute recorded call transcribed with Mini ($0.021) and followed by a 5,000-character spoken response ($0.065) totals about $0.086 for those two model calls. This excludes the reasoning model, any managed voice-agent charge, retries, storage, telephony, and application infrastructure.

Do not mix standalone models with the managed agent rate

Mistral separately lists Voice Agent API pricing at $0.04 per minute with Mistral Small 4 for reasoning, or $0.08 per minute with a larger reasoning model. That is a managed, end-to-end pipeline price. Treat it as its own product meter; do not add individual transcription and TTS line items on top unless your contract or usage records show they are billed separately. Conversely, a self-built pipeline should count its transcription, reasoning, and speech-generation calls independently.

Choose by the workload you have

Batch Mini is the starting point for recorded calls, meetings, and support archives. Move to Small when an evaluation set shows its transcript quality reduces human correction or downstream errors enough to offset the extra $0.0005 per minute. Use realtime transcription when text must arrive during the conversation; its $0.008/min rate is about 2.3 times Mini batch, so reserve it for latency-sensitive workflows.

For FinOps, log audio duration, generated character count, selected model, and final task outcome together. Report cost per accepted minute, resolved call, or completed interaction. The raw rate card is low, but retries, idle live sessions, repeated speech, and reasoning can dominate the realized cost. All figures above are Mistral's published API prices checked October 27, 2026; verify the live rate card and your billing configuration before forecasting.

Sources

• Mistral Docs: October 20 release changelog
• Mistral: Voxtral 3 announcement and pricing table
• Mini Transcribe 3 model card
• Small Transcribe 3 model card
• Mistral Docs: audio model selection
• Mistral Docs: managed voice agents

Related

• Cost per successful task
• Model pricing and routing budgets
• Budgeting for agent workloads

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com