LAST UPDATED · 16 AUGUST 2026
Attic Standard Methodology
How the AI Pricing Index is built, and what it does not claim
Attic Standard publishes an independent price benchmark for AI inference. Every figure is a price a vendor published, captured on a date, held with the original string it was read from. Nothing is estimated, interpolated or inferred. This page states how the numbers are constructed and where the limits are.
PART ONE
Principles
What the benchmark is, and what it refuses to do
A benchmark is only as good as the worst price in it. Attic Standard indexes published prices only, and states plainly when a reading rests on a narrow basket.
§ 1.1
Four rules
These govern every decision in the pipeline. Where they conflict with a tidier chart, the chart loses.
PUBLISHED PRICES ONLY
If a vendor does not publish a rate, it does not enter the index. Nothing is estimated, interpolated or carried forward. Where a price is captured on a basis that cannot be converted without an assumption, the SKU is reported and held outside the basket rather than converted.
DETERMINISTIC EXTRACTION
Prices are read by rules written against each vendor's published format, not by a model interpreting a page. The same inputs always produce the same output, and the original string is stored beside the normalised value so any figure can be traced to what the page said.
THE VENDOR THAT SETS THE PRICE
A price is indexed once, against whoever sets it. Resellers passing through another vendor's rate are excluded, since counting both would double the supply and understate concentration.
DISCLOSED LIMITS
Basket size, vendor count and coverage are published beside every reading. Where a week's movement rests on a small number of models, the chart says so.
§ 1.2
Verification
Every SKU carries a last-verified date recording when its price was last checked against the vendor's page. Extraction runs weekly against every tracked vendor, and each run reports what it read, what it could not read, and what appeared on the page that is not yet tracked. A model that vanishes from a vendor's page leaves the basket rather than holding its last known price.
Historical audit. In August 2026 every visible step in every published series was walked back to the individual SKUs that produced it, and each one was either confirmed as vendor repricing or corrected at source. Eight corrections were made across 81 price records, each one a real number attached to the wrong model, unit or week rather than an invented figure. The corrected history is what the site now publishes.
PART TWO
Composition
What each index contains
§ 2.1
How the indexes cut the market
Each index answers one buyer question. Every index page carries its own definition, constituent counts and coverage; this is the shape of the set.
MODALITY
What does this kind of inference cost? Text, multimodal, image, audio, video, voice and embedding, each priced in the unit its market actually trades in. These partition the tracked universe.
CHANNEL
Where should I buy it? Model developers, cloud marketplaces, inference platforms and neoclouds. These partition the token universe, so no SKU sits in two of them.
TIER
What does capability cost? Frontier, mid-tier, budget and specialised, assigned to the model rather than the listing so a model carries the same tier wherever it is sold.
LICENCE
What does openness save? Open weights, open weights with restrictions, and proprietary. The restricted set is a subset of open weights, so the two must not be summed.
ORIGIN
Where was it built? Origin describes where a model was created, not where it is sold, so this cut and the supply map differ substantially.
CAPABILITY
Reasoning is recorded as a capability rather than a tier, so a reasoning model also appears in its tier index. This cut overlaps the others and must not be summed with them.
Partitions can be summed; cuts cannot. Each index page states which it is.
§ 2.2
What is excluded
These are excluded because they are not comparable to a published on-demand rate, not because they do not matter.
EXCLUDED
REASON
Negotiated and enterprise rates
Not publicly verifiable, so not reproducible by a reader
Committed-use discounts
Buy a commitment rather than a unit of inference
Volume tiers above the first
Only the base tier is indexed, so a reader sees the rate available without commitment
Batch and asynchronous rates
Price a different service level
Subscriptions and credit bundles
Cannot be attributed to a unit of inference
Free tiers and trials
Not a production cost
Resold capacity
The price is set by the underlying vendor and is already indexed there
Delisted models
Leave the basket in the week they stop being published
PART THREE
Construction
From a published rate to an index level
§ 3.1
Normalisation
Vendors publish in different units. A price is converted only where the conversion is arithmetic and requires no assumption about how the service is used.
PUBLISHED AS
INDEXED AS
CONVERSION
per 1M tokens
per 1,000 tokens
Divide by 1,000
per token
per 1,000 tokens
Multiply by 1,000
per character
per 1,000 characters
Multiply by 1,000
per hour of audio
per minute
Divide by 60
per second of audio
per minute
Multiply by 60
per compute second
Not converted
Depends on hardware profile
per megapixel
Not converted
Depends on output resolution
per credit or bundle
Not converted
Depends on bundle composition
The unit map holds 51 published forms across six modalities. The conversions above are those a reader is most likely to meet; the full mapping is available on request. Where a form cannot be converted, the SKU is tracked, reported in the index coverage note, and held outside the basket. That is why several indexes cover well under half the SKUs in their modality, and why the excluded count is published rather than buried.
Currency. Almost all tracked vendors publish in US dollars. Where a vendor publishes in another currency, the converted figure is not treated as equivalent to a published dollar rate: it is indexed only where the vendor also publishes a dollar price, and the conversion is recorded as a change in basis rather than a change in price.
§ 3.2
The chained matched-model index
The problem the construction solves is that the market grows every week. Between December 2025 and August 2026 the tracked universe went from 740 priced SKUs to over 5,500. A naive average would move every week on composition alone.
A SKU must be priced in two consecutive weeks before it can move an index. Each week's change is measured only across SKUs present in both weeks, and that change is applied to the previous week's level. A model arriving or leaving therefore changes what the basket contains without producing a price movement, and enters the calculation from its second observation.
Index_t = Index_(t−1) × geometric mean over matched SKUs of ( price_t ÷ price_(t−1) )
Every model counts once. The weekly change is the geometric mean of each matched SKU's own price change, so a model priced at $120 per million tokens moves the index no more than one at $0.02. Averaging the prices instead would weight each model by what it costs, which on a basket spanning a six-thousandfold price range means a handful of expensive models decide the reading. This construction is the Jevons index, the elementary aggregate formula used by national statistical offices, including in the UK, US and EU consumer price indices, for exactly this situation: a basket with no reliable quantity weights.
No volume weighting. Vendors do not publish inference volumes, and any weighting Attic Standard invented would be a guess presented as a measurement. Each model therefore carries equal weight. The index measures how listed prices moved, not what the market spent.
Directions are indexed separately. Most vendors price prompt and completion independently and the ratio between them varies by an order of magnitude, so combining them would average two different markets. Cached input is indexed over the models that publish a cached rate, measured against their own input, so the cached series is not comparable in level to the full index.
§ 3.3
Benchmark and spot
Two numbers are published for each index because they answer different questions.
BENCHMARK
The chained level, rebased to January 2026 = 100. It measures movement on identical models, with composition removed by construction. It does not quote a price and is not meant to: a level of 97.2 says listed prices are 2.8 percent below where they started the year.
SPOT
The median price across every qualifying SKU live this week, composition included. It quotes what a typical model actually costs today, alongside the range from the 25th to the 75th percentile. It moves when the mix of models changes, which the benchmark deliberately does not.
The gap between them carries information. Benchmark falling while spot holds means existing models are being repriced downward; spot falling while benchmark holds means cheaper models are entering the market. Reported alone, either one would mislead.
§ 3.4
Index rebalancing
A dotted vertical line on a chart marks a week where a small number of models produced most of the movement.
In an equally weighted index every member carries the same influence, so a week in which one model reprices sharply can move the level as much as a week in which many move a little. Both are real, but they mean different things to a buyer, and the chart alone cannot tell them apart.
A series is marked in a week where the level moved one percent or more and a single matched model accounted for half or more of the total 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 not applied by hand.
A mark describes that week alone. It carries no implication about the weeks around it, and it does not mean the reading is doubtful: every price behind a marked step has been verified against the vendor's published page. Marks are most common on cached input, where the population of models publishing a cached rate is smallest.
§ 3.5
Safeguards
SAFEGUARD
PURPOSE
Per-week change bound
The matched-set change is capped each week. Movement beyond the bound is recorded and flagged rather than compounded into the chain. Under the geometric construction this is a backstop rather than a load-bearing control, since a single extreme member no longer dominates.
Two-observation rule
A SKU enters the calculation from its second observation. A model seen once cannot yet have changed price, so it cannot contribute movement.
Positive prices only
A zero or absent price removes the SKU from the basket for that week rather than being treated as free.
Year-to-date anchor
Year-to-date change is measured against the raw median of the underlying SKUs at the start of the year rather than against the chained level, anchoring it to an observable market value.
§ 3.6
Revision policy
Prices are restated only where the published figure can be shown to be wrong and the correct figure is observable. Where a value is wrong but the correct one was never captured, the reading stands and the limitation is recorded rather than a number being invented. Restatements are applied across the affected history and the series is recomputed in full.
August 2026 revision. The weekly change was previously computed as the ratio of the matched basket's average price between two weeks. On a basket spanning a six-thousandfold price range this weighted each model by its price level, and in one measured week a single model at $120 per million tokens produced 97 percent of a 698-model index move. The construction was changed to the geometric mean of individual price changes and the full history recomputed. Cumulative index levels moved by less than half a point; weekly readings are materially more accurate. The change is documented here rather than applied silently.
PART FOUR
Limits
What this benchmark does not measure
§ 4.1
Out of scope
These are deliberate exclusions. A price index that also claimed to measure quality or speed would be doing neither well.
NOT MEASURED
WHY
Performance and latency
Requires running the models, which is a different discipline with different infrastructure. A cheaper model is not necessarily a slower one, and this index says nothing either way.
Output quality
Attic Standard records published specifications but does not score models. Capability rankings change faster than prices and would make the index a judgement rather than a measurement.
Total cost of ownership
Egress, storage, support and integration vary by deployment and are not published per unit. The index measures the listed rate, which is one input to that calculation rather than a substitute for it.
What buyers actually pay
Negotiated and committed rates are not public. The index measures listed prices, which set the reference point most contracts are negotiated against.
Volumes and market share
No vendor publishes inference volumes. The index measures price, not spend, and does not weight by either.
§ 4.2
Known limitations
LIMITATION
EFFECT
Coverage gaps in early history
Some series ran on a small number of vendors in their first weeks, because capture had not yet reached the full market. Basket size and vendor count are published for every week, and affected steps carry a rebalancing mark.
Unconvertible units
Several modalities exclude a substantial share of tracked SKUs because their pricing basis cannot be converted without an assumption. The excluded count and the bases involved are published in each index's coverage note.
Vendor-side errors
A vendor occasionally publishes a wrong price. Where this is detectable against the model's own history or a sibling listing it is corrected; where it is not, it is captured faithfully and the index reflects what was published.