> ## Documentation Index > Fetch the complete documentation index at: https://docs.compute-desk.com/llms.txt > Use this file to discover all available pages before exploring further. # Token Prices: Metadata > LLM inference prices per model, in USD per 1M tokens. Returns the values you can filter this data by, what it covers, and how far back you are able to read it, in `history_starts_at`. This is the best first call for any data type: it tells you what a query will accept before you write one. It is not paged. ## Example ```bash theme={null} curl -H "Authorization: Bearer $TOKEN" \ "https://data-api.compute-index.com/v2/llm/token/prices/metadata?canonical_model_id=anthropic/claude-opus-4-8" ``` ## OpenAPI ````yaml pricing/openapi.json GET /v2/llm/token/prices/metadata openapi: 3.1.0 info: title: GPU Pricing API (v2) description: > Pricing data for GPU compute — rental and purchase prices, settled transactions, published indices, and LLM inference prices. Every data type exposes the same four endpoints over one filter vocabulary, one response envelope and one error format, so learning one is learning all of them. **Start at `/metadata`**, which lists the values a type accepts and the range it holds. <details> <summary><strong>Endpoints</strong> — what the four paths return</summary> | path | returns | | --- | --- | | `/latest` | the current state, one row per thing | | `/as-of` | the state as it stood on a given date | | `/history` | rows over a time window | | `/metadata` | the type's own vocabulary — the valid filter values, and what it covers | Some types have fewer. Settled transactions are events, and "the current state of events" is not a question, so those types serve `/history` and `/metadata` and nothing else. </details> <details> <summary><strong>Responses and paging</strong> — the envelope, and how to get page two</summary> Every response is `{"data": [...], "meta": {...}}`. `data` is the rows; `meta` describes the request they answer — the sort applied, the window resolved, and the paging position. Nulls are always present. A null means the value is not published, which is different from zero and different from absent. `page_size` sets the rows per page (default 100, max 1000). Follow `meta.next` for the following page, and stop when it is null. `meta.next` is an object describing the following page three ways: `url` is an absolute URL ready to call, and `path` plus `query_params` are the same request split up for clients that build their own. Use whichever suits your HTTP layer. Each carries a `page_token`, and the token IS the request — it replays your filters, window and sort exactly — so a following page cannot silently change the question. There is nothing to merge in and nothing to add: do not construct or edit a token, and send no parameters beside it. There is no offset, and no total count. </details> <details> <summary><strong>Filtering, sorting and time</strong> — anchors, sort fields, UTC windows</summary> Filters are per type and listed on each operation. Values are exact matches unless the parameter says otherwise, and where a list is accepted it is comma-separated. **Some types require an anchor** — at least one filter that narrows the query, such as a vendor or a GPU. Asking for everything at once is rejected with a clear error naming the fields that qualify, and `/metadata` lists them. `sort_by` and `sort_order` control ordering, from that type's own set of sortable fields; `meta.sort` reports what was applied. All timestamps are UTC, ISO-8601. Windows are `start`/`end` and are half-open — `start` is included, `end` is not — so consecutive windows tile without double-counting a boundary row. </details> <details> <summary><strong>Access and errors</strong> — tokens, entitlements, rate limits</summary> Send a bearer token: `Authorization: Bearer <token>`. Entitlements are per data type, so a token reaches the types it is granted and 403s on the rest. Rate limits are reported on every response in the `X-RateLimit-*` headers. Errors are RFC 9457 problem details, served as `application/problem+json` with `type`, `title`, `status` and `detail`. </details> contact: name: API Support email: david@compute-index.com version: 2.0.0 servers: - url: https://data-api.compute-index.com security: [] tags: - name: GPU Rental Prices description: >- What it costs to rent GPU compute, as listed by the vendors themselves. One row is one vendor's price for one product. Prices are normalised so they compare across vendors: read `price_usd_per_gpu_hour`. `raw_price` and `raw_price_unit` keep whatever the vendor actually published, for when you need the original. A price that changes produces a new row rather than editing the old one, so `/history` reads as a list of changes and `/as-of` as a single state. Collection queries need an anchor — a vendor, a GPU, or another narrowing filter. `/metadata` lists which fields qualify here. - name: GPU Rental Aggregated Transactions description: >- Settled GPU deals published as blends rather than individually. Every row is a volume-weighted blend of at least two deals, so no single deal can be identified from it. `price` is the only price figure on a row: there is no median, range or spread to ask for. **Do not sum across rows.** Some rows share a deal with the row before them, so adding up `total_gpu_count` or `total_gpu_hours` will double-count. `is_rolling` marks those rows, and `blend_axis` names what was combined to make up the row. There is no `/latest` or `/as-of` — use `/history` with a window. - name: GPU Price Indexes description: >- Published GPU price indices — blended transaction indices at daily and 30-minute resolution, by GPU family or by GPU, per region — and the GPU reference price, the fleet-weighted mean of the per-GPU indices. **One index per request.** Every data endpoint takes a required `id`; `/metadata` lists the ids you hold, and an id you do not hold returns 404. **Two row shapes, told apart by `type`.** A `daily` row's `timestamp` is a calendar date (`2026-09-02`); an `intraday` row's is the UTC instant its interval starts (`2026-09-02T10:00:00+00:00`). `/metadata` gives each index's `row_type`, so you know the shape before you ask. Sort by `timestamp` or `value`. `/metadata` says who computes each index. `p5` and `p95` bracket `value` where an envelope is published, and are null where it is not. - name: GPU Hardware Prices description: >- What it costs to buy GPU hardware outright, from resale, marketplace and reseller listings — the capex counterpart to rental prices. One row is one listing on one day it was seen. A listing that was not seen on a given day has no row for that day, so `/latest` returns each listing's most recent observation rather than a single snapshot. Check `last_observed` in `/metadata` before reading a product's price as current. **Seller identity is not published** — no platform, seller or listing URL, and none of them filterable. What you get instead is the classification a price needs to be read correctly: `market`, `channel`, `price_type`, `source_type`, `confidence` and `counterparty_class`. Check the `includes_*` fields too — a whole-chassis price and a bare-module price are not the same number. - name: LLM Token Prices description: >- Published prices for LLM inference, normalised to USD per 1M tokens so they compare across sources that quote per-token, per-1k and per-1M. **One model has many rows.** A row is identified by the whole combination of source, model, serving provider, price type, region, metric and qualifiers — input and output are priced separately, and the same model appears once per source and once per provider. Filtering to a model and expecting a single row will not work. Implausible prices are flagged rather than removed: `is_price_outlier` marks them and they are included by default, so a response is the data as recorded. Pass `exclude_outliers=true` to drop them. paths: /v2/llm/token/prices/metadata: get: tags: - LLM Token Prices summary: 'LLM Token Prices: filters and coverage' description: >- LLM inference prices per model, in USD per 1M tokens. Returns the values you can filter this data by, what it covers, and how far back you are able to read it, in `history_starts_at`. This is the best first call for any data type: it tells you what a query will accept before you write one. It is not paged. operationId: metadata_llm_token_prices_metadata_get responses: '200': description: Successful Response content: application/json: schema: $ref: '#/components/schemas/MetadataEnvelope_TokenPricesMetadata_' '401': description: Unauthorized — `code` is `unauthenticated`. content: application/problem+json: example: status: 401 title: Unauthorized detail: Missing or invalid bearer token. code: unauthenticated request_id: 01JQ2X8N4T6V9WZ0ABCD '403': description: Forbidden — `code` is `insufficient_entitlement`. content: application/problem+json: example: status: 403 title: Forbidden detail: The token does not carry the scopes this endpoint requires. code: insufficient_entitlement request_id: 01JQ2X8N4T6V9WZ0ABCD '404': description: Not Found — `code` is `not_found`. content: application/problem+json: example: status: 404 title: Not Found detail: No such resource, or none you are entitled to see. code: not_found request_id: 01JQ2X8N4T6V9WZ0ABCD '422': description: Unprocessable Content — `code` is `validation_failed`. content: application/problem+json: example: status: 422 title: Unprocessable Content detail: A parameter failed validation. code: validation_failed request_id: 01JQ2X8N4T6V9WZ0ABCD '429': description: Too Many Requests — `code` is `rate_limited`. content: application/problem+json: example: status: 429 title: Too Many Requests detail: Rate limit exceeded; see the X-RateLimit-* headers. code: rate_limited request_id: 01JQ2X8N4T6V9WZ0ABCD security: - HTTPBearer: [] components: schemas: MetadataEnvelope_TokenPricesMetadata_: properties: data_type: type: string title: Data Type description: The data type this vocabulary describes. examples: - prices history_starts_at: anyOf: - type: string - type: 'null' title: History Starts At description: >- The caller's licensed history floor for this type, as an ISO date. Null means unrestricted — either the caller has no floor, or this type is a published series that is not floor-bound. Reported here so a client can discover its own entitlement in one call rather than by overshooting a window and reading the clamp. examples: - '2020-01-01' data: $ref: '#/components/schemas/TokenPricesMetadata' additionalProperties: false type: object required: - data_type - history_starts_at - data title: MetadataEnvelope[TokenPricesMetadata] TokenPricesMetadata: properties: anchors: items: type: string type: array title: Anchors description: At least one is required on every collection query. price_unit: type: string title: Price Unit description: What `price_per_mtok` is denominated in. examples: - usd_per_mtok unit_family: type: string title: Unit Family description: >- The quantity priced, stated because it bounds what this type IS: the same table holds per-image, per-second and per-request lines and they are not served here. examples: - tokens models: items: $ref: '#/components/schemas/TokenModel' type: array title: Models description: >- The anchor catalogue. `lab` is a refinement, not an anchor — this is how you find the ids. vocabularies: $ref: '#/components/schemas/TokenVocabularies' additionalProperties: false type: object required: - anchors - price_unit - unit_family - models - vocabularies title: TokenPricesMetadata description: >- The `token-prices` vocabulary. Lists currently priced models only. A model not priced since March is not queryable state to hand a client — `/history` and `/as-of` still reach its past. TokenModel: properties: canonical_model_id: type: string title: Canonical Model Id examples: - anthropic/claude-opus-4-8 lab: anyOf: - type: string - type: 'null' title: Lab examples: - anthropic modality: anyOf: - type: string - type: 'null' title: Modality examples: - text context_length: anyOf: - type: integer - type: 'null' title: Context Length examples: - 200000 priced_metrics: type: integer title: Priced Metrics description: How many distinct metrics carry a current price. examples: - 2 sources: type: integer title: Sources description: How many distinct sources price it. examples: - 1 additionalProperties: false type: object required: - canonical_model_id - lab - modality - context_length - priced_metrics - sources title: TokenModel description: One currently-priced model in the anchor catalogue. TokenVocabularies: properties: metric: items: type: string type: array title: Metric source: items: type: string type: array title: Source price_type: items: type: string type: array title: Price Type provider: items: type: string type: array title: Provider modality: items: type: string type: array title: Modality lab: items: type: string type: array title: Lab region: items: type: string type: array title: Region additionalProperties: false type: object required: - metric - source - price_type - provider - modality - lab - region title: TokenVocabularies description: >- The value sets actually present in the current data. Derived on request rather than fixed in advance, so what is listed here is always what a filter on that field can currently match. securitySchemes: HTTPBearer: type: http scheme: bearer ````