diff --git a/docs/getting-started/enterprise-api.md b/docs/getting-started/enterprise-api.md index 2dbc483..0e2c4d9 100644 --- a/docs/getting-started/enterprise-api.md +++ b/docs/getting-started/enterprise-api.md @@ -176,7 +176,7 @@ Rows are newest first. Page with `cursor` until `next_cursor` is `null`; treat t |-------|---------| | `kind` | `chat` — a token-priced call you can re-derive from `input_tokens`/`output_tokens` and the [published rate](../products/intelligence/pricing.md) as adjusted by [what a call costs](#what-a-call-costs). `service` — a per-call or per-unit price (search, media, RPC) that only the gateway holds. | | `cost_state` | `priced` — final. `free` — nothing to charge (a catalogue read, a free model). `pending` — usage recorded, charge not yet settled; it will change, do not treat it as $0. | -| `service_tier` | The processing tier the upstream reported for this call. `flex` is billed at **half** the standard rate, so this is what explains a half-price line against the published rate. Absent means standard — as it is on every row before 2026-09-23 and for every provider that reports no tier. See [What a call costs](#what-a-call-costs). | +| `service_tier` | The processing tier the upstream reported for this call. `flex` is billed at **half** the standard rate and `fast` (or `priority`) at **twice** it, so this is what explains a line at half or double the published rate. The doubling applies only when the call asked for Fast; a `priority` row on a call that did not ask is billed at standard. Absent means standard — as it is on every row before 2026-09-23 and for every provider that reports no tier. See [What a call costs](#what-a-call-costs). | | `job_id`, `job_status` | Media only. A `202` create and every later poll of the same job share one `job_id`; the charge lands on the poll that completed. Fold rows by `job_id` to get one line per generation. The job is **one call**, made by its create: each poll is a row — poll as often as you like — but not a call. Sum `cost_usd` over every row; count calls skipping the polls. | | `usage_unit`, `usage_units` | For non-token charges: what was billed (`image`, `second`, `character` …) and how many. `null` on token-priced rows. | | `billing_basis`, `image_input_tokens` | `token` on an OpenAI image row (`openai/gpt-image-2`, `gpt-image-2.5-*`): a `service` by path, but one you can rebuild — `input_tokens`/`output_tokens` at OpenAI's published per-million rates, `image_input_tokens` being the slice of input at the image-input rate. Absent on every other row. See [Images from OpenAI are billed by token](#images-from-openai-are-billed-by-token). | @@ -203,7 +203,7 @@ If your tooling already reads OpenRouter's billing API, point it at `https://api **There is no per-call fee and no minimum charge on this host.** The x402 transaction fee and the minimum payment exist to make on-chain micropayments viable; a key-metered account settles against credit, not a chain, so both are zero here. You pay metered usage times the rate, and nothing is added to it. -Two things move the rate away from the flat number on the catalogue. Both are visible on the ledger row, so any line you cannot reconcile against the published rate is explained by one of them. +Three things move the rate away from the flat number on the catalogue. All three are visible on the ledger row, so any line you cannot reconcile against the published rate is explained by one of them. ### Long context @@ -224,12 +224,21 @@ Four things worth knowing before you build on it: - **Availability is the model provider's decision, and it is a limited beta.** OpenAI publishes the current list on the flex tab of their own pricing page; a model that does not offer Flex either rejects `service_tier` outright or serves the call at standard rates and bills accordingly. A model whose `openai/` id is served through a partner pool — the `-pro` tiers — is billed at standard rates whatever tier ran upstream, because the discount is not one we receive and so not one we can pass on. - **`service_tier` on every `GET /v1/usage` row is how you confirm it.** `flex` on the row means that line was billed at half. Absent means standard, which is also the case for every row before 2026-09-23 and for every provider that reports no tier at all. +### Fast + +`openai/gpt-6-luna` can run in OpenAI's Fast mode: up to 2.5x faster, at **twice** the standard rate in every column. That is $0.20 input and $1.00 output per million tokens, $0.40 / $1.50 past 272K, cached read $0.02 and cached write $0.25. Send `service_tier: "fast"` (`"priority"` is accepted as the same thing), or call the alias `openai/gpt-6-luna-fast`, which `GET /v1/models` lists at the Fast prices. Usage is recorded under `openai/gpt-6-luna` either way. + +- **We bill from the tier the response reports.** If OpenAI serves the call at standard speed (the response says `default`), you pay the standard rate. +- **Only `openai/gpt-6-luna` offers it.** `service_tier: "fast"` on any other model answers `400` with `fast_tier_unavailable` before anything is charged, and so does Fast on a background Response. +- **Not on an absolute agreed price.** An account with an agreed per-million price on `openai/gpt-6-luna` gets `400` with `fast_tier_not_contracted`; ask us to add Fast to your terms. A percentage discount does not block it: see below. +- **Only on this host.** blockrun.ai and sol.blockrun.ai do not take `service_tier: "fast"` today. + ### If your account has negotiated terms -The two tiers interact with a contract differently, and deliberately so: +Long context, Flex and Fast interact with a contract differently, and deliberately so: -- An **absolute agreed price** per million **suppresses** both tiers. The long-context and Flex rates are not part of what was agreed, so layering them on either charges more or discounts more than the contract says. -- A **percentage discount** **follows** both tiers. A percentage is off whatever you would otherwise pay, and the tier is part of what you would otherwise pay. +- An **absolute agreed price** per million **suppresses** long context and Flex. Their rates are not part of what was agreed, so layering them on either charges more or discounts more than the contract says. Fast is **refused** instead (`400`, `fast_tier_not_contracted`): the agreed price was set against standard serving, and Fast costs us twice that. +- A **percentage discount** **follows** all three. A percentage is off whatever you would otherwise pay, and the tier is part of what you would otherwise pay, so a Fast call pays twice the rate and then gets the discount. ## Images from OpenAI are billed by token diff --git a/docs/resources/changelog.md b/docs/resources/changelog.md index c73019d..2fb73c3 100644 --- a/docs/resources/changelog.md +++ b/docs/resources/changelog.md @@ -7,6 +7,13 @@ description: All notable changes to BlockRun — gateway endpoints, model lineup All notable changes to BlockRun, newest first — gateway endpoints, model lineup, pricing, and SDK releases. +## [2026-11-18] + +### Added — Fast mode for GPT-6 Luna on api.blockrun.ai +- `openai/gpt-6-luna` with `service_tier: "fast"`, or the alias `openai/gpt-6-luna-fast`, runs in OpenAI's Fast mode on [api.blockrun.ai](../getting-started/enterprise-api.md#fast): up to 2.5x faster at twice the standard rate ($0.20/M in, $1.00/M out; $0.40/$1.50 past 272K). Billed at the standard rate when OpenAI serves it at standard speed. Not on other models, background Responses or an absolute agreed price (a percentage discount still applies), and not on blockrun.ai or sol.blockrun.ai. + +--- + ## [2026-11-17] ### Added — GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Hydrogen