tiyuvta Get an API key

Pricing

Two ways to buy the same tokens: per-token with a hard p95, or a whole replica reserved. The numbers are being finalized and are not published yet.

fig. 01 — how the meter works (live today)
  • Worker-truth counts. The usage block on every response — streams included — carries token counts from the worker that did the work, not a client-side estimate.
  • Aborts billed to the abort point. Disconnect mid-stream and you are billed for what was generated up to the abort, not for the full request.
  • Cached tokens itemized. prompt_tokens_details.cached_tokens appears on every response, so the cache discount is auditable per request.
serve-compat receipts → research/serve-compat-20260802
fig. 02 — the structure, ahead of the numbers
pricing not published yet

Prices ship here and in the /v1/models metadata at the same moment — gateways cross-check it programmatically, and so can you. Until then, request a key and current terms are shared directly.

plan input / 1M tokens output / 1M tokens cached input / 1M tokens
interactive per-token · hard p95, never preempted — — —
dedicated a whole replica, reserved by the hour — / hr per reserved replica, tokens not metered
cached input discounted and itemized per request in prompt_tokens_details.cached_tokens

What is already decided, independent of the numbers: interactive is per-token with an admission-controlled p95 and is never preempted; dedicated is a whole replica of the SKU with no neighbors of any kind and no per-token meter; cached input is discounted and itemized on every response. There is no perpetual free tier; new accounts get a small one-time evaluation credit.

Get an API key Read the docs

OpenAI SDK compatible — change one line · evaluation credits included