Predictions Api

Integrator calls POST /v1/predictions with body {version: 'a1b2c3...', input: {...}} for an SDXL prediction. A teammate proposes switching to the model-slug form POST /v1/models/stability-ai/sdxl/predictions to 'always get the la…

Pin version id (immutable) for production traffic so behavior is reproducible. Use the model-slug form only for exploratory work or when you have an explicit auto-upgrade policy — latest_version moves on every upstream push and can silently change input schema, output shape, or safety behavior.

Integrator wants a synchronous response from a fast FLUX-schnell prediction and sends header Prefer: wait=30.

Use Prefer: wait=<seconds> (up to ~60s) for short predictions where a single HTTP response is simpler than polling. If the prediction does not finish within the wait, Replicate returns the prediction in a non-terminal state and the client must fall back to polling /v1/predictions/{id} or webhooks —…

Async prediction created with status=starting. The integrator polls GET /v1/predictions/{id} every 200ms until terminal.

Use exponential backoff (e.g., 1s → 2s → 4s, cap at 10s) and key off status transitions starting → processing → succeeded|failed|canceled. Prefer webhooks for long predictions to eliminate polling cost. Honor any Retry-After header on 429.