A 403 from OpenAI's transcription API on an unenabled model costs nothing, so fallback is free

Repro:

1. Use an OpenAI project that has not been granted access to `gpt-4o-mini-transcribe`. 2. Call `audio.transcriptions.create(model="gpt-4o-mini-transcribe", file=...)` with any audio file. 3. The call returns a 403 permission error. Zero usage is billed for that request.

The rejection is an authorization check, not an inference failure. It happens before any inference runs, so there is no audio processing to charge for. That matters for how you structure model selection.

### The pattern this allows

If you want the cheaper or newer transcription model but cannot be sure every project or key you deploy to has it enabled, you do not need a separate capability probe, a config flag, or a list of allowed models per project. Try the preferred model first, catch the permission error, and retry with a model you know is enabled.

1. Call the preferred model. 2. On a 403 permission error, call the fallback model with the same file. 3. On any other error (rate limit, 5xx, bad audio), handle it as you normally would. Do not treat those as a signal to fall back.

The failure path costs one round trip and no billed usage. The success path costs the same as if you had called the preferred model directly.

### Caveats

- Catch only the permission error. A blanket `except` that falls back on every failure will hide real problems, such as a malformed file, and may bill you twice for a request that failed after inference started. - You pay the latency of one extra request on the failure path. If the project will never have access, cache the result for the process lifetime so you only take that hit once. - I have only verified this for the 403 on a model the project cannot access. I have not measured what other rejections do to billing, so I would not assume the same behavior for them.

### Why it is worth writing down

A lot of harness code defends against a cost that does not exist here. Pre-flight checks and per-project model tables add state that can drift from what the account actually allows. The API already answers the question for free, at the moment you need it, and it is always current.