generated: '2026-09-17'
method: derived
source: >-
  openapi/discountedtokens-api-openapi.json + https://discountedtokens.com/llms.txt +
  live probes of https://discountedtokens.com/v1 (2026-09-17).
summary: >-
  Cross-cutting semantics for the DiscountedTokens reseller API. It is a thin OpenAI/Anthropic-
  compatible pass-through: three inference write operations plus a model-list read, one prepaid
  Bearer key, and an OpenAI-style error envelope. No idempotency, pagination, expansion, or
  request-tracing conventions are documented, and no reversal surface exists.
authentication:
  styles: [bearer_token, api_key_header]
  bearer: 'Authorization: Bearer <prepaid-key>  (OpenAI + Responses compatible)'
  api_key_header: 'x-api-key: <prepaid-key>  (accepted on /messages, Anthropic-compatible)'
  see: authentication/discountedtokens-api-authentication.yml
client_integration:
  note: >-
    No first-party SDK is published (absent from npm and PyPI, 2026-09-17). The provider is designed
    to be called with the stock OpenAI, OpenAI Responses, or Anthropic clients by overriding the
    base URL to https://discountedtokens.com/v1. This is the documented integration path (llms.txt).
pagination:
  style: none
  note: /models returns a single unpaginated { object:"list", data:[...] } of 3 models; inference calls are not list surfaces.
streaming:
  supported: true
  note: createChatCompletion accepts a `stream` boolean (OpenAI-compatible SSE streaming); observed in the request schema.
versioning:
  style: unversioned-path
  path_segment: /v1
  spec_declared: 1.0.0
  note: The /v1 segment has not moved; info.version reads 1.0.0. See lifecycle/.
error_envelope:
  media_type: application/json
  shape: '{ "error": { "message": string } }'
  see: errors/discountedtokens-api-problem-types.yml
rate_limit_signaling:
  documented: false
  observed_headers: none
  note: >-
    No X-RateLimit-*/RateLimit-*/Retry-After headers were returned on live unauthenticated responses
    (Cloudflare-fronted). See rate-limits/. Billing, not a request quota, is the documented governor.
request_tracing:
  documented: false
  note: No provider request-id header is documented; responses carry only Cloudflare's cf-ray.
idempotency:
  documented: false
  header: null
  coverage: none
  note: >-
    No Idempotency-Key mechanism is documented and none appears in the OpenAPI. The mutating surface
    (/chat/completions, /responses, /messages) is billed per successful request with no replay
    primitive, so a client that retries a timed-out inference call has no safe-replay guarantee and
    may be billed twice. No `Idempotency` pointer is emitted — this is a genuine zero, not a missing
    pointer.
reversibility:
  applicable: true
  grade: none
  reversal_operations: []
  note: >-
    The write surface is LLM inference (generation) plus prepaid billing. No reversal operation
    (cancel/refund/void/undo) is exposed in the API and no refund window is documented; the terms
    describe successful requests as billed. An agent cannot un-send or refund a completion through
    the API. Recorded as an honest `none` (a write surface with no reversal path), not `na`.
dry_run_mode:
  supported: false
  note: No dry-run/simulation mode is documented; /v1/models (a free read of prices) is the only way to preview cost before a billed call.