Specification files

Status: cheapgpu.endpoint.v1-draft.4. This draft can change before a stable release. No production API endpoint or supplier eligibility is provided by publishing these files.

Download contracts

The manifest lists every artifact and its status. JSON Schema files use draft 2020-12. Their $id URLs identify the intended documentation location; local tests resolve them from files without network access.

Artifact	Format	Purpose
Common definitions	JSON Schema	IDs, revisions, profile and result scope
Model catalog	JSON Schema	Shared identity, typed tasks, native interface references and measurement units
Source catalog	JSON	Four registered OpenAI snapshot identities; supplier qualification remains separate
Endpoint configuration	JSON Schema	Immutable stored configuration, not a registration request
Chat text profile	JSON Schema	Request, completion, chunk, usage and error definitions
Responses profile	JSON Schema	Native input/output items, functions, reasoning, structured output, inline images, typed events and usage
Conformance run	JSON Schema	Immutable scope, policies, budgets and deadline
Conformance report	JSON Schema	Terminal result and bounded-run metadata
Conformance suite	JSON	Required cases and assertion descriptions
Stored Responses lifecycle	JSON	Native stored defaults, retrieval, pagination, continuation and deletion cases
Background Responses lifecycle	JSON	Background creation, polling, cancellation and cursor-based stream recovery cases
Responses field options	JSON	Nullable function schemas, reasoning/text options and distinct JSON strictness cases
Responses constraints	JSON	Field constraints, schema limits and trusted base/fine-tuned applicability
Responses suite	JSON	Responses-specific required cases, independent of Chat

Examples include a catalog, configuration, check run, request, completion, stream, and report. The files under examples/ are synthetic. Their language/video limits and check limits are not production defaults or Seedance facts. The separate source catalog contains official-documentation facts; it is not a synthetic fixture or a live service.

Public API availability

See Public API for the availability of API key authenticated endpoints. The downloads above describe the supplier protocol draft.

Versioning

Revision 4 adds the native Responses draft. The manifest’s profiles array identifies each profile’s schema and suite; it has no implicit default profile. Configuration and report scopes require the capability set for the selected profile. The shared run limit max_output_tokens_per_request bounds the native output-token ceiling, including reasoning. The old draft shape is not accepted. Chat’s body, stream and required cases retain their existing profile/suite IDs. Manifest, schemas, source records and synthetic examples advance together; there is no legacy parser or cross-profile translation. Typed task metadata still does not define or implement video inference operations.

The Responses reference links its JSON, function continuation, structured output, image and SSE examples. They are synthetic protocol examples for offline validation, not actual checker results.

Record the contract, profile, configuration, catalog, suite, and limit-policy versions separately. A new draft revision must update the manifest, schemas, examples, and tests together. Before external use, freeze immutable artifact URLs for each released version and define its compatibility window.

An incompatible body, streaming, or required-capability change needs a new profile ID. A changed required-case set needs a new suite version and a recheck policy. Never silently treat an old passing report as a pass for a new suite.

Release blockers

Before suppliers can join, provide public HTTPS endpoint registration with secure credential provisioning, a fixed model catalog and production checker limits. V1 does not require a separate endpoint-control verification step. Define result authenticity, eligibility expiry and operational health policy, capacity coordination, and secure gateway egress. Production network egress must enforce the endpoint security requirements.

Complete the selected native profile’s contracts and implementations, including its capability and usage semantics. Responses has a defined draft and offline examples; its runtime, network checker and real supplier acceptance remain delivery gaps. Stateful Responses, hosted tools, broader schemas/image inputs, Anthropic Messages and video task contracts remain incomplete. The existing text-only profile is not a final product restriction.

Before buyer traffic, implement endpoint inference and streaming against backend admission/dispatch, then define trusted usage acceptance, successful collection, partial/cancelled/filtered results, currencies, precision, commission, rounding, and deadline values. None of these financial values is selected by the examples. No per-request manual review or platform-funded supplier income guarantee is part of the design.

The Responses schema suite and recursive examples cover local references, recursive output and nested anyOf in the draft profile.

The constraint examples pair valid and invalid values for the nine string formats, patterns, numeric bounds and exact multiples, and array item counts. They are synthetic protocol vectors, not supplier results.