Deferred processing

Hydracept deferred processing uses OpenAI Flex (service_tier=flex) at 50% of standard token rates.

It applies to:

Interactive invoke and stream stay on standard processing. Image, audio, video, and 3D jobs are unchanged — those providers do not expose this discount today.

When it applies

PathProcessing
POST /v1/capabilities/{key}/jobs (job_async)Deferred when the routed model supports it. HTTP 429 retries stay on Flex; a Flex 400 retries once on standard rates.
POST /v1/inference/pinned and /pinned/bulkStandard by default. Opt in with processing: "deferred" on eligible OpenAI Responses pins. Same pin, one logical model execution, no truncation rewrite, no Flex→Standard fallback. Flex-capacity 429s retry inside the admission deadline and are recorded on the receipt. Omitted execution timeouts on Flex pins use the 600s platform max.
POST /v1/capabilities/{key}/invokeStandard
StreamStandard

Discover support on the capability descriptor:


GET /v1/capabilities/text.reasoning.high.v1

Look for:


"features": {
  "deferredProcessing": {
    "available": true,
    "discountBps": 5000,
    "appliesTo": ["job_async", "pinned"],
    "defaultFor": ["job_async"],
    "optInFor": ["pinned"]
  }
}

Receipts record processingTier and deferredDiscountBps on every pin. The discounted upstream basis is recorded only when deferred actually ran.

Pinned native envelopes add service_tier: "flex" only when the caller opts in (plus the existing store: false). Hydracept does not send Responses truncation and does not change max_output_tokens.

BYOK and managed

The 50% applies to the provider token basis, not the Hydracept service fee.

See Billing, Durable jobs, and Pinned Execution.