Transcribe audio (OpenAI wire format)

Served on the terms a completion is: the key's chat scope and policy, the account's credits with a hold at the worst case, and one metered usage row settled at what the upstream measured. A model with no published price, or one priced in a unit the endpoint does not report, is refused on house money with 503 pricing_unavailable and served on your own provider key. With no provider configured the answer is 503 upstream_unavailable. No fallback chain and no x402: a media request names one model and needs a key. The upload is capped at ANTIMATTER_AUDIO_MAX_BYTES (4000000 bytes by default, under Vercel's request ceiling; at most 26214400), answered 413 request_too_large past it. whisper-1 is charged per second of audio, gpt-4o-transcribe and gpt-4o-mini-transcribe per token of each kind, and an OpenRouter model the cost its answer reports. The hold reserves the file at 16 kbps, the longest it plausibly plays. srt and vtt are refused: they carry no usage, so their cost cannot be measured; verbose_json carries the timings.