Skip to main content
POST
Speech-to-Text
gpt-transcribe transcribes uploaded recordings. Use the dedicated multipart route POST /api/v2/stt; an audio_url passed to /generate does not invoke this route. GET /api/v2/models lists the STT models available to the account.

File and language

  • audio: a nonempty file up to 25 MiB (25 * 1024 * 1024 bytes).
  • GPT-Transcribe formats: MP3, MPGA, MPEG, WAV, M4A, OGG, WEBM, MP4 and FLAC.
  • model: gpt-transcribe; if omitted, the server selects the currently available OpenAI STT model.
  • language: auto or one two-letter code such as en or ru.
The server measures duration and validates the recording before charging. Practically silent audio is rejected. The duration-based price starts at 1 token; the quote returns the exact amount. Speaker labels, word timestamps, subtitles and additional context controls are outside this GPT-Transcribe file mode.

Quote before execution

POST /api/v2/stt/quote takes the same audio, model and language fields. It does not call the model or debit tokens. The response includes model, duration_seconds, token_cost and an opaque quote_fingerprint. Submit the quoted values as expected_token_cost and quote_fingerprint with the paid request. If the file, model, controls or price changed, the server returns 409 before generation and charging. Obtain a new quote and confirm it again; do not automatically replay the paid POST.

Response

id identifies the historical generation; the transcript is in result_text. A direct request without quote fields remains valid: the server validates the file and calculates the current price before execution. Use quoting for explicit price confirmation. After catalog cutover, whisper and gpt-4o-transcribe are unavailable for new selections. Historical identities, prices and results remain; an explicit retired-model request is not silently redirected. Check other STT models against their current contracts; passing an extra form field does not make it supported.