Integration notes

• Always send the public model name, never upstream branch names.
• For video send seconds (mind each model's discrete/fixed/range values); prefer tier over size strings.
• Use reference_images for image/video-to-X (mind counts and whether required — see each model page).
• Put model-specific params inside metadata (e.g. omni's first_last_frame / source_video_url); pass them via extra_body when using the OpenAI SDK.
• Read the result from output_url in the query response (present once complete) — don't guess upstream fields.
• Tasks are async — always poll and finish only on a terminal status. ~3–5s for images, 5–10s for video.
• Reference images accept https URLs or base64 data URLs (data: prefix handled automatically).