Attic Standard Methodology

A census, not a scrape. Every capture compares what a source lists against what we track. New models are surfaced for review, a model missing from its source leaves the basket that week, and a listing is retired only when its absence is confirmed.

Dry run before write. Every capture runs first without writing. Anything our rules cannot read with certainty is blocked and resolved before a single price is recorded.

Continuous audit. We audit prices, units, vendors and model identities against their sources on a regular cycle. Where a published figure proves wrong and the correct value is observable, we restate the history and recompute every affected index in full.

Identity checks. Two standing checks guard model identity. Prices are tested against each model's own market: a listing more than three times its model's median across vendors is flagged and confirmed at source. And no vendor may carry two prices for the same model and leg, so a serving path, a context band or a mode priced differently is identified and handled under a stated convention rather than averaged into the model.

Channel and license indexes divide listings without overlap; tier, origin and use case indexes overlap the others. We never add one index level to another.

Admission, model by model. No model enters an index automatically. We review each one and admit or exclude it under a registered rule, with the reason and its source recorded. A decision is made once for the model and applies wherever it is sold, and a newly listed model is held out until it has been reviewed.

What admission prevents. Serving paths, quantized copies, regional variants, context and quality bands, and endpoints of a model already represented are excluded, so no model is counted twice. Where a vendor moves a model name to a newer version, the name is excluded and the versions are indexed as the separate products they are. A superseded generation stays in the modality indexes while it is sold, and a model built on another organization's base is attributed to that base where the question requires it.

Price conventions. Where a model is offered at several resolutions, quality settings, context lengths or times of day, the index carries a stated convention: the lowest published band, 720p video without audio where a silent rate is published, and the peak rate. Each convention is applied identically across every vendor.

Every conversion is stored with the price it was applied to. Where a form cannot be converted without an assumption, we track the listing and hold it outside the basket.

Currency. Almost all tracked vendors publish in US dollars. Where a vendor publishes in another currency, we do not treat the converted figure as equivalent to a published dollar rate: it is indexed only where the vendor also publishes a dollar price, and a change of currency basis is recorded as a change in basis rather than a change in price.

Every model counts once. A model sold by many vendors carries one vote: its listings' changes are combined into a single change for that model before models are combined. Both steps use the geometric mean, so a model priced at $120 per million tokens moves the index no more than one at $0.02. This is the Jevons index, the elementary aggregate formula recommended by the international Consumer Price Index Manual for baskets without reliable quantity weights, applied here at both stages.

No invented weights. Vendors do not publish inference volumes, so each model carries equal weight, and the index measures how listed prices moved rather than what the market spent.

Directions are indexed separately. Most vendors price input and output independently, and the ratio between them varies by an order of magnitude. Cached input is indexed over the caching cohort, the models that publish a cached rate, and the cohort's own input and output are shown beside it so cached prices are read against the same models.

The gap between them carries information. A falling benchmark with a steady spot means existing models are being repriced; a falling spot with a steady benchmark means cheaper models are entering, or coverage has widened to include them.

A series is marked in a week where its level moved 1 percent or more and one matched model accounted for at least half of that movement. Both figures come from the index calculation itself, so a mark can never disagree with the published number. The test runs on every index every week and is never applied by hand.

A mark describes that week alone. We use the term concentrated repricing rather than index rebalancing, because rebalancing means a change to the basket, which never moves this benchmark.

Each model counts once. Where a measure describes models, prices are reduced to a median per company and then to a median per model, so a model sold by many companies carries the same weight as one sold by a single company. Every price relationship is read within one model: output against its own input, cached against its own standard rate, one seller against another on the same model.

Repricing follows the listing. Repricing measures follow each listing, one company's price for one model in one unit. A move under 0.5 percent is not a change, a move reversed within two weeks is treated as temporary and excluded together with its reversal, and a listing's legs moving together count as one change.

Publication and base. A measure is published only in weeks when its population reaches ten. Each is read against its May 2026 average, with changes stated in percentage points for shares and in the multiple itself for ratios, and the full history is recomputed at every publication.