Transcription models on OpenRouter

Every speech-to-text model OpenRouter listed on the day of the snapshot, with its price per hour of audio. The table is generated from data/stt-models.json, which admin/build/snapshot-models.mjs writes from GET /api/v1/models?output_modalities=transcription and each model's /endpoints. The build never fetches: the page says what the file says, and the file says when it was taken.

How to read it

• Per hour is the listed prompt price times 3,600. OpenRouter's guide says duration-priced models are "priced per second of audio input"; the anchor that confirms the unit is openai/whisper-1 at $0.0001 a second, which is OpenAI's published $0.006 a minute.
• Token-priced models (a non-zero completion price) bill by audio and text tokens; their hourly cost depends on the speech and is not computed here. Read usage.cost from a real request.
• Unit unclear: two Azure models list prices (0.1 and 0.36) that cannot be per second (that would be $360 and $1,296 an hour), and the unit is not stated. Not computed.
• Range is across the providers serving the model. OpenRouter does not apply provider ordering (order, only, ignore) to transcription requests, so which provider, and so which price, serves a given request is not yours to choose. The extension shows usage.cost, never this table.
• A meeting with both channels and VAD is billed for roughly the minutes anybody spoke, so an hour of meeting is close to an hour of billed audio, not two.

Snapshot taken 2026-11-03 00:24 UTC: 29 models.

The bake-off candidates

Phase 0 tests six of these on real recordings in English, Portuguese and both: openai/whisper-large-v3, openai/whisper-large-v3-turbo, mistralai/voxtral-mini-transcribe-3-0, nvidia/nemotron-3.5-transcribe, elevenlabs/scribe-v2, openai/gpt-transcribe. The choice is made on quality first; this table only says what each would cost if it wins.