generated: '2026-09-10'
method: searched
source: https://modelrush.ai/docs (api-keys, concepts/errors, concepts/webhooks,
  operations/observability) + openapi/modelrush-public.openapi.yaml + live probe of
  api.modelrush.ai/v1/models
authentication:
  style: bearer-api-key
  header: 'Authorization: Bearer $MODELRUSH_API_KEY'
  notes: Keys are server-side only; keys in query strings are not accepted. See
    authentication/modelrush-authentication.yml.
idempotency:
  coverage: none
  note: No Idempotency-Key or equivalent client replay-protection header is
    documented on any write operation. The only idempotency guidance published is
    consumer-side -- webhook handlers must be idempotent because delivery is
    at-least-once, deduplicated via the x-modelrush-event-id header. Retrying a
    billable POST (chat, image, video, speech) can therefore double-bill; the docs
    mitigate with "a small maximum attempt count" rather than replay keys.
pagination:
  style: none
  note: 'List responses (GET /models, GET /webhooks/endpoints) return complete
    list envelopes (object: list, data: [...]) with no cursor or page parameters.'
request_tracing:
  headers:
  - x-request-id
  - x-modelrush-registry-revision
  note: Observed on a live GET /v1/models response 2026-09-10. Docs instruct
    capturing the ModelRush request/generation ID and correlating with application
    traces via the Requests dashboard.
versioning:
  path_prefix: /v1
  registry: dated registry revisions (x-modelrush-registry-revision, e.g. 2026-09-08.1)
  changelog: https://modelrush.ai/docs/changelog
error_envelope:
  shape: '{error: {code, message, type}}'
  format: custom (not RFC 9457)
  catalog: errors/modelrush-problem-types.yml
rate_limit_signaling:
  status: 429
  headers:
  - Retry-After
  note: No X-RateLimit-* quota headers documented; see rate-limits/modelrush-rate-limits.yml.
streaming:
  style: SSE
  note: POST /chat/completions returns text/event-stream when stream:true, with
    OpenAI-compatible stream_options.
async_jobs:
  pattern: submit -> poll /predictions/:id (or /videos/generations/:id) -> or
    receive signed webhook on terminal state; poll_url is returned on the job
    object. Fallback polling remains available when webhooks are delayed.
regions:
  note: Requests may pin an execution region via the region request field; model
    and region availability is time-sensitive -- call GET /models and GET /regions
    before selecting.
reversibility:
  summary: partial
  surfaces:
  - surface: asynchronous media generation (Predictions)
    reversal: cancelPrediction
    grade: documented
    note: DELETE /predictions/{id} requests cancellation and returns the terminal
      Prediction state; a 409 is returned when the job is already terminal, so the
      window is "before the Prediction reaches a terminal state". The docs do not
      state a time-bound window, so this grades documented rather than verified.
  - surface: media uploads
    reversal: deleteUpload
    grade: documented
    note: DELETE /uploads/{id} removes a private upload; uploads also carry an
      expires_at and age out on their own. No restore window is stated.
  - surface: synchronous billable generations (createChatCompletion,
      createImageGeneration, createImageEdit, createSpeech, createTranscription)
    reversal: none
    grade: none
    note: Completed generations are billed and cannot be reversed; no refund or
      void operation is published.
  - surface: webhook endpoints
    reversal: none-documented
    grade: none
    note: The OpenAPI publishes list/create only; no delete/disable endpoint
      operation is in the public contract (an enabled flag exists on the object).