Get started

Tape

Source

Every current price in one class, flattened to one row per priced cell.

text
GET /api/v1/markets/{class}/tape
bash
curl https://vitna.ai/api/v1/markets/model/tape

Path parameters

ParameterValues
classmodel, tooling, retrieval, compute, extraction, speech, translation, image, video, music

Anything else answers 404 unknown_class.

Response

json
{
  "class": "model",
  "tape": [
    {
      "sku": "gpt-5-pro",
      "name": "gpt-5-pro",
      "provider": "OpenAI",
      "market": "First-party",
      "featured": true,
      "dimension": "input_per_mtok",
      "dimensionLabel": "Input price",
      "amountUsd": 15,
      "unit": "USD per 1M tokens",
      "since": "2026-08-07T00:00:00.000Z",
      "confirmedAt": "2026-08-23T08:00:11.402Z",
      "fetchedAt": "2026-08-23T08:00:11.402Z",
      "stale": false,
      "derived": false,
      "sourceUrl": "https://openai.com/api/pricing/",
      "evidenceSha256": "9f2c..."
    }
  ],
  "meta": { "generatedAt": "...", "source": "database", "count": 918, "units": ["USD per 1M tokens"] }
}

Fields

FieldMeaning
skuSlug, for /skus/{slug}/history
providerWho sells it
dimensionWhat it is metered by
amountUsdThe price, in unit
sinceWhen this meter's price started, in this band: the since of its open period in /skus/{slug}/history, not the oldest date across the sku
confirmedAtWhen it was last confirmed
fetchedAtWhen the document was read
stalePast its source's cadence
derivedComputed by Vitna rather than read from the page
derivationHow, when derived is true
sourceUrlThe page a human should read
evidenceSha256The archived document
observationIdThe leaf this figure is, for its proof

Model rows additionally carry market and featured. A derived row carries derivation in every class. Compute rows carry the key on every row, as null where the figure was read as printed; every other class leaves it out. A model row served from the committed snapshot (meta.source: "snapshot") holds no read per figure, so a derived figure there says that it was derived and not how.

A row priced in a context band also carries bandFromTokens, bandToTokens and bandLabel.

The leaf behind a figure

observationId names the observation this figure is, which is what turns a number on the board into something you can prove is in the signed record:

bash
curl "https://vitna.ai/api/v1/observations/<observationId>/proof"

It is the observation that last confirmed the price, which is the same read confirmedAt and evidenceSha256 describe. All three on one row therefore refer to one read, so checking the document and checking the leaf are checks of the same thing. fetchedAt, stale, sourceUrl and derived are that read's too.

That holds per row, not per sku. One model's figures can come from different reads: OpenAI's and Anthropic's flex and priority rates come from a second source, the tiers page, read in a fetch of its own, and a table a page stops printing keeps its periods open on the last read that saw it while the rest of the row is confirmed every day. Each figure carries its own read, so a fresh input price is never called stale because its batch rate is, and a batch rate never cites a document it is not in.

One id per row of the tape, not per sku. A single fetch writes one observation per cell it reads, so a model priced on four meters has four leaves even though one document covers them all. That is why the tape carries the id and the board does not.

Ids are strings. They are bigserial, and a JSON number stops being exact above 2^53.

Context bands

Some providers price one model differently by how much context a request carries. A price therefore applies in a band rather than universally, and a sku priced that way publishes one entry per band per meter.

FieldMeaning
bandFromTokensExclusive lower bound
bandToTokensUpper bound, null when unbounded
bandLabelWhat the provider calls it

So a model split at 272,000 tokens publishes an input price for (0, 272000] and another for (272000, null], each naming its own band.

A band is not a dimension and not a sku. It is a property of the price, which is why it travels on the row rather than multiplying either registry.

One row per cell

A model publishing input, cached input and output prices produces three rows. This is the flat shape on purpose: it is joinable and countable, where the board shape the site renders is not.

Empty classes answer 200 with an empty tape

A registered class with nothing priced returns "tape": [], which is the honest answer and the same one the site gives. A 404 would say the class does not exist, which is different and untrue.

An empty tape means one of two things, and `/markets` says which: a class Vitna does not read yet (wired: false), or a class it reads whose sellers price nothing today (wired: true). video is the second kind since Google shut down Veo 3.1 Preview on 2026-10-22, which took every figure the class held.

meta.units

The distinct units present, which is more than one for most classes.

The speech tape carries two units across three meters, since transcription is billed by the audio-minute pre-recorded and again streaming, and synthesis by the million characters. The retrieval and extraction tapes are single-unit. The tooling tape is metered four ways. The model tape carries two, because a model that emits images is billed by the image rather than by the token.

An extraction row carries one field no other tape has, page, which says which kind of page its figure is for. Two per-page rates are not comparable without it: the cheapest is a page read for its characters and the dearest is a page read against a schema.

Caching

public, max-age=60, stale-while-revalidate=3600.