Transcriptions

POST /v1/audio/transcriptions

Auth: routing key (pllm_…). Speech-to-text that behaves like OpenAI’s transcription API, served through the Codex login connected to the key’s account - no separate transcription bill. The key must have a connected Codex seat.

Point any OpenAI SDK at https://api.proxyllm.ai/v1 and call audio.transcriptions.create as usual.

Request body

Send multipart/form-data, the same shape an OpenAI SDK sends.

Field	Type	Notes
file	file, required	The audio to transcribe (mp3, mp4/m4a, wav, webm, …).
model	string	gpt-4o-transcribe (default when omitted is gpt-4o-mini-transcribe). whisper-1 is not available on a Codex login and is served by gpt-4o-transcribe instead - see below.
language	string	ISO-639-1 hint, forwarded.
prompt	string	Optional context to bias the transcript, forwarded.
response_format	string	json (default) or text. Other formats are normalized to json.
temperature	number	Forwarded.

A JSON body is also accepted as a convenience for hand testing: {"file": "<base64>", "filename": "voice.mp4", "model": "gpt-4o-transcribe"}.

curl https://api.proxyllm.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer pllm_your_routing_key" \
  -F model="gpt-4o-transcribe" \
  -F file="@voice.mp4"

from openai import OpenAI

client = OpenAI(base_url="https://api.proxyllm.ai/v1", api_key="pllm_your_routing_key")
client.audio.transcriptions.create(model="gpt-4o-transcribe", file=open("voice.mp4", "rb"))

Models

Two transcription models run on a Codex login:

• gpt-4o-transcribe
• gpt-4o-mini-transcribe (the default when model is omitted)

whisper-1 is blocked for a Codex login, so a request for it is served by gpt-4o-transcribe and the response header x-proxyllm-model-substituted: whisper-1 names what you asked for. x-proxyllm-model always names the model that actually ran.

Response

response_format=json (the default) returns the OpenAI transcription shape:

{
  "text": "No, I've done it before.",
  "usage": {
    "type": "tokens",
    "input_tokens": 20,
    "output_tokens": 9,
    "total_tokens": 29
  }
}

response_format=text returns the transcript as text/plain.

Response headers report which lane served the request and what it would have cost at API rates (a Codex seat bills $0): x-proxyllm-provider, x-proxyllm-model, x-proxyllm-model-substituted, x-proxyllm-cost-usd, x-proxyllm-api-equivalent-usd, x-proxyllm-latency-ms.

Errors

Status	Code	When
400	no_audio_file	No audio file was found in the request.
401	signup_required	Missing key, or a raw provider key was sent.
401	invalid_api_key	Unknown or revoked routing key.
401	codex_reconnect_required	The connected Codex login expired; reconnect it.
402	budget_exceeded	Monthly budget exhausted.
405	(none)	Method other than POST.
502	no_codex_session	No Codex login is connected for this key.
502	capacity_reached	The Codex account is out of transcription capacity for now.
502	upstream_error	The transcription service was unavailable; try again.

Full list: error codes.