generated: '2026-07-21'
method: searched
source: >-
  https://platform.stepfun.com/docs — cross-cutting request/response semantics
  gathered from the quickstart, API reference (chat completions, Messages,
  Responses, token count), error-codes, prompt-cache guide, and pricing/rate
  limit pages.
description: >-
  How the StepFun open-platform API behaves across operations: OpenAI-compatible
  REST surface (plus an Anthropic-compatible Messages API), Bearer API-key
  authentication, SSE streaming, prompt caching with discounted cache-hit
  pricing, token-based usage accounting, an OpenAI-style error envelope, and
  top-up-tiered rate limits.

base_url: https://api.stepfun.com/v1
api_style: REST over HTTPS, JSON requests/responses, OpenAI-compatible; WebSocket for realtime audio

compatibility:
  openai: >-
    The core surface (chat/completions, models, files, images, audio,
    responses) follows the OpenAI API shapes; the docs' quickstart officially
    instructs reuse of the OpenAI SDK with base_url https://api.stepfun.com/v1.
  anthropic: >-
    POST /v1/messages is compatible with the Anthropic Messages API format;
    the Anthropic SDK works with base_url https://api.stepfun.com.
  step_plan: >-
    Step Plan subscription traffic uses the parallel prefix
    https://api.stepfun.com/step_plan/v1/... with the same shapes.

authentication:
  scheme: Bearer API key (Authorization header)
  key_console: https://platform.stepfun.com/interface-key
  docs: https://platform.stepfun.com/docs/zh/quickstart/overview
  detail: authentication/stepfun-authentication.yml

idempotency:
  supported: false
  notes: No idempotency-key mechanism is documented.

pagination:
  style: openai-list
  notes: >-
    List endpoints (files, vector stores, voices) return OpenAI-style list
    objects; the docs do not document a cross-cutting cursor contract.

streaming:
  supported: true
  mechanism: 'stream: true returns Server-Sent Events chunks (OpenAI delta format); reasoning models stream thinking deltas'
  realtime: >-
    Bidirectional realtime voice over WebSocket at wss://api.stepfun.com/v1/realtime
    (OpenAI Realtime-style session/conversation/response events) and streaming
    TTS at wss://api.stepfun.com/v1/realtime/audio.
  detail: asyncapi/stepfun-realtime-asyncapi.yml

prompt_caching:
  supported: true
  notes: >-
    Automatic prompt caching with separate, discounted cache-hit input pricing
    (e.g. step-3.7-flash ¥1.35/M tokens uncached vs ¥0.27/M cached).
  docs: https://platform.stepfun.com/docs/zh/guides/developer/prompt-cache

usage_accounting:
  mechanism: >-
    Responses carry usage.prompt_tokens, usage.completion_tokens and
    usage.total_tokens; billing is token-metered per model. A dedicated
    POST /v1/token/count endpoint pre-counts tokens.
  docs: https://platform.stepfun.com/docs/zh/guides/pricing/intro

versioning:
  scheme: uri-path
  current: v1
  notes: All endpoints live under /v1 (or /step_plan/v1); no date-based versioning documented.

error_envelope:
  media_type: application/json
  rfc9457: false
  shape: '{ "error": { "message", "type" } }'
  detail: errors/stepfun-error-codes.yml
  docs: https://platform.stepfun.com/docs/zh/api-reference/error-codes

rate_limits:
  signal_status: 429
  model: cumulative top-up tiers V0-V5 (concurrency / RPM / TPM)
  detail: rate-limits/stepfun-rate-limits.yml
  docs: https://platform.stepfun.com/docs/zh/guides/pricing/details

other_conventions:
  - name: Content moderation
    detail: HTTP 451 is returned when request or response content fails moderation review.
  - name: Balance gating
    detail: HTTP 402 is returned when the account balance is insufficient; GET /v1/accounts reports balance.
  - name: Model deprecation
    detail: Models are retired on announced dates with named replacement model IDs; see lifecycle/stepfun-lifecycle.yml.