OpenAI (the reference)¶
The moving target. This page exists to document what OpenAI itself does, not because we expect anyone else to match it perfectly.
Surface¶
OpenAI implements the full canonical surface plus several extensions that haven't propagated to OSS servers:
/v1/responses— newer than chat completions, structured around responses (multi-turn, multi-output) instead of message arrays./v1/audio/speechstreaming./v1/threads,/v1/assistants,/v1/files— the Assistants API./v1/batches— async batch processing at half price./v1/realtime— WebSocket-based bidirectional streaming.
The catalog tracks the first three as ext. Assistants, batches, and
realtime are intentionally out of scope — they're shaped
differently enough that probing them as "OpenAI-compatible" against an
OSS server doesn't tell you anything useful.
Auth¶
Authorization: Bearer <key> is required. The prober sends the key
from OPENAI_API_KEY (or AIOC_API_KEY).
OpenAI-Beta headers gate features that are in flight. The prober
doesn't add them — we want to compare against the stable surface.
Probing OpenAI itself¶
Why probe the reference? To detect spec drift in our own catalog.
If OpenAI changes the chat completions response shape and the prober
flags WARN against api.openai.com, the catalog is out of date,
not OpenAI.
Things to keep in mind¶
- Rate limits. Even with
aioc's ≤ 2 requests / endpoint budget, a free-tier key may hit the per-minute cap on the chat probe. Use a paid tier or expect SKIP rows. - Cost.
max_tokens=4keeps each chat call to ~¢0.001 ongpt-4o-mini. Image generation is ~¢4 per call (512x512, n=1ongpt-image-1). The full probe is roughly $0.05 against the cheapest tier. - Sandbox. OpenAI does not run in
sandbox=truemode by default. The probe sends real prompts; review the catalog's bodies if you're concerned about content.