POST /api/v2/generate.
Uploaded-file transcription uses the separate POST /api/v2/stt/quote and POST /api/v2/stt routes.
GET /api/v2/models returns the models currently available to the account.
Text-to-Speech
gpt-realtime-2.1-mini creates a complete speech file, including Russian speech.
It offers ten voices, speed 0.25–1.5, voice instructions and six formats.
Text up to 4096 characters costs 3 / 8 / 15 / 20 tokens for the inclusive
500 / 1500 / 3000 / 4096-character buckets, with no additional debit after generation.
The default voice is Alloy.
Send model and the text in prompt; put voice, speed, instructions and response_format in params.
The asynchronous response contains an id: poll GET /api/v2/generate/{id} until completed,
then download through GET /api/v2/download/{id} with Authorization.
See the example and full parameters.
After catalog cutover, this model replaces openai-tts, openai-tts-hd and gpt-4o-mini-tts
for new selections. Historical results and prices remain. Other TTS models, including
MiniMax and ElevenLabs, retain their own controls and current quotes.
Interactive Realtime voice sessions belong to a later stage.
Speech-to-Text
gpt-transcribe accepts an uploaded file up to 25 MiB and language=auto or one two-letter code.
The server measures duration; the recording-based price starts at 1 token.
First send the same file to /stt/quote without a debit. After accepting the quote, submit
the file, expected_token_cost and quote_fingerprint to /stt. A changed quote returns
409 before a paid call. Read the transcript from result_text.
See formats and a working example.
After catalog cutover, whisper and gpt-4o-transcribe are unavailable for new selections;
historical results are not renamed. Other STT models follow the current catalog.

