Speech-to-text API pricing: Whisper, GPT-Transcribe, Scribe, Voxtral and the cheap open models

Updated Nov 7, 2026

Speech-to-text models are sold by the minute of audio, so the question is simple: what does an hour of transcription cost? The catalog has 33 speech recognition models from 20 providers; 42 offers are priced per minute, the rest per token (they are in the model tables but not in the per-minute list below). Prices from the 7 November 2026 collection; an hour = 60 × the price per minute.

Price per hour of audio

• Open models at the cheapest hosts - Qwen3 ASR 0.6B and Nemotron 3.5 ASR about $0.012 an hour, Whisper Large v3 Turbo from $0.012 (typically $0.03), Qwen3 ASR 1.7B $0.03, Parakeet TDT 0.6B v3 $0.08.
• Mid-range commercial models - StepAudio 2.5 ASR about $0.06, Grok STT and MAI-Transcribe-2 about $0.10, ElevenLabs Scribe v2 $0.11, Soniox stt-rt-v6 $0.12, NVIDIA Nemotron 3.5 Transcribe 30B $0.12, Voxtral Mini Transcribe 3 $0.21, Voxtral Small $0.24.
• OpenAI - GPT-Transcribe $0.27 an hour, Whisper (whisper-1) $0.36 at OpenAI and from $0.03 at other hosts of the open Whisper.
• Real time - GPT-Live-Transcribe and GPT Realtime Whisper about $1.02 an hour: streaming with low delay costs several times more.

Whisper Large v3 Turbo: every provider

Live data: this table is built from today's prices.

Whisper (whisper-1): every provider

Live data: this table is built from today's prices.

How to choose

• Batch transcription of recordings: an open model (Whisper Large v3 Turbo, Qwen3 ASR) at a cheap host costs cents per hour - test it on your own audio, accents and noise first.
• Quality and extras out of the box (speaker labels, punctuation, many languages): GPT-Transcribe, Scribe v2, Voxtral, Soniox.
• Live captions and voice agents: a streaming model - priced higher, check the delay.
• Your own volume: the images, video and voice calculator prices any number of minutes at every model; how to use it.

All speech recognition offers: the catalog, speech to text.