How the reference price for AI inference is built, verified and maintained.
Published
November 9, 2026
Covers
Principles, construction, measures and governance
For
Analysts, auditors and counsel reviewing the indexes
Principles
Attic Standard sells, routes and brokers no inference and holds no position in the prices it measures. Six rules govern every decision, from capture to publication.
IndependenceThe instrument, not a participant. No commercial relationship bears on an assessment, and no vendor sees a figure before publication.
Published prices onlyIf a vendor does not publish a rate, it does not enter an index. Nothing is estimated, interpolated or modeled.
Deterministic extractionPrices are read by fixed rules written for each source, so the same inputs always give the same output.
The vendor that sets the priceEach price is indexed at the vendor that sets it. Routers and aggregators that pass a rate through are excluded.
Recorded decisionsEvery model carries an admission decision with its rule, reason and source, so a reader sees why it is in or out.
Full disclosureBasket size, vendor count, start date and base size sit beside every reading, and a week moved by one model is marked.
Capture and verification
Prices come from the source each vendor maintains, and nothing is published until every tracked vendor is verified.
1Capture at source
From pricing and model APIs, catalogs, rate cards and pricing pages, each price kept with its source, date and original string.
2Dry run
Every capture runs first without writing. Anything the rules cannot read with certainty is resolved before a price is recorded.
3Verify
Index membership is reconciled with the prices behind it, and any material move is traced to its models and confirmed at source.
4Publish
A week is released only when every tracked vendor has been captured and verified for it.
A census, not a scrapeEach capture compares what a source lists with what is tracked. New models go to review, and a missing model leaves the basket that week.
Outlier checkA listing more than three times its model’s median across vendors is confirmed at source before it counts.
One price per model and legA vendor cannot carry two prices for one model and leg. Serving paths, context bands and modes follow a stated convention.
Continuous auditPrices, units, vendors and identities are audited against source. A proven error is restated and every affected index recomputed.
Composition
Each index answers one buyer question, and its members must be substitutes for one another, sold in one unit.
ModalityWhat does this kind of inference cost?Text, multimodal, image, video, speech, voice and embeddings, each in the unit its market trades in
ChannelWhere should I buy it?Model developers, cloud marketplaces, inference platforms and neoclouds
TierWhat does a model of this standing cost?Flagship, core and compact, by each maker’s own lineup, not by benchmark scores
LicenseWhat does openness cost or save?Published, restricted and unpublished weights
OriginWhere was it built?Where a model was created, not where it is sold
Use caseWhat does work-specific inference cost?Models sold for reasoning and models built for coding
Channel and license indexes divide listings without overlap, while tier, origin and use case overlap the others, so index levels are never added together.
Admission
One identity per modelEvery listing resolves to one canonical model, so a model sold under many names is priced and counted once.
Admission, model by modelNo model enters automatically. Each is admitted or excluded under a registered rule, and a new listing waits for review.
No double countingServing paths, quantized copies, regional variants and moving aliases are excluded. Versions are indexed as separate products.
What is excluded
Reason
Excluded
Why
Not public
Negotiated, enterprise and committed rates
Cannot be verified by anyone
Not standard
Higher volume tiers, batch and promotional rates
Not the rate open to any buyer
Not a unit
Subscriptions, credits, free tiers, GPU time
Cannot be compared across vendors
Already counted
Routers, aggregators, duplicates and variants
Indexed at the vendor that sets it
Not one product
Price bands, priority paths, moving aliases
A band of one price, or a mix of models
Not a substitute
Task-specific tools and delisted models
Outside the question, or no longer sold
Price conventions. Where a model is sold at several resolutions, quality settings, context lengths or times of day, the index takes the lowest published band, 720p video without audio where a silent rate exists, and the peak rate. Where a vendor lists one model under two names, the lower price is captured.
Construction
The market grows every week, so the index level measures change only on the same models from one week to the next.
Units
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 megapixel, per clip
as published
None; each its own market
per compute second, per credit
not converted
Depends on hardware or bundle
A price is converted only where the arithmetic needs no assumption about use, and the conversion is stored with it.
Currency. Every price is captured in the currency its vendor publishes. Dollar prices build the global indexes, and euro and yuan prices build their own regional indexes at native prices, never converted. A vendor pricing in a currency with no published index yet, such as yen, won or rupee, is tracked and held outside the baskets, and a change of currency basis is recorded as a change in basis, not in price.
The index level
Model change = geometric mean of ( pricet ÷ pricet−1 ) across the model’s matched listingsIndext= Indext−1 × geometric mean of model changes across matched models
One vote per modelListings are combined into one change per model before models are combined, so a model priced at $0.12 per 1,000 tokens moves the index no more than one at $0.00002.
No invented weightsVendors publish no volumes, so every model weighs the same. This is the Jevons index the international CPI Manual recommends without quantity weights.
Each direction on its ownInput, cached input and output are indexed separately. Cached input covers the models that publish a cached rate, read beside their own input and output.
Index level and spot price
Index levelThe chained level, May 2026 = 100 (regional indexes October 19, 2026 = 100), on identical models. A level of 97.2 means listed prices are 2.8% below their May average.
Spot priceWhat the market charges now: the median across models of each model’s median vendor price, with the 25th and 75th percentiles. A new listing counts from its first priced week.
Reading the gap. A falling level beside a steady spot means models are being repriced; a falling spot beside a steady level means cheaper models are entering. Where a basket averaged fewer than 15 models in May, or has changed by more than half, the card shows its base count, as in 64 models (39 at base).
Concentrated repricing. A dotted line marks a week in which the level moved 1% or more and one model produced at least half of it. Both figures come from the calculation itself, the test runs on every index, and it is never applied by hand.
Safeguards
Safeguard
What it does
Complete capture
No index moves because a vendor was missed.
Two observations
A listing moves the level only from its second consecutive priced week.
Minimum constituents
Eight models for input and output, five for cached input, in every week from the first May 2026 week.
Change bound
A weekly change is bounded at 50% either way, and a week that reaches it is recorded.
Latest verified price
Where a price changes twice in a capture week, only the last verified price counts.
Positive prices only
A zero or missing price leaves the basket for that week; it is never read as free.
Market measures
Nine readings of how the market prices, and four of how each index behaves, built from the same admitted prices as the indexes.
One value per modelPrices reduce to a median per vendor, then per model, and every price relationship is read within one model.
A lasting changeA move under 0.5% is not a change, and a move reversed within two weeks is excluded with its reversal.
Measure
What it reads
Price structure
Output Premium
Share of token models whose output rate is more than 4× their input rate
Caching Discount
Median saving of cached input against the same model’s input rate
Caching Availability
Share of token models sold with a cached rate by at least one vendor
Price dynamics
Repricing Activity
Share of listings with a lasting change in 13 weeks, and the share that were cuts
Repricing Depth
Median size of the lasting price changes in the last 13 weeks
Post-Launch Drift
Share of models launched 4 to 13 weeks ago now priced below their launch week
Competition
Multi-Vendor Spread
Median ratio of highest to lowest input price, on models with two or more vendors
First-Party Premium
Share of models priced higher by their maker than by the median independent host
Marketplace Premium
Share of models priced higher on a cloud marketplace than by the median independent host
Index behavior, on each card
Price Dispersion
Gap between the 25th and 75th percentile spot, as a multiple of the median
Vendor Concentration
Share of the basket’s listings held by its largest vendor
Repricing Incidence
Share of basket listings with a lasting change over 12 published weeks
Index Volatility
Standard deviation of the weekly change over the last 12 weeks, in points
A measure is published only when its population reaches ten, read against its May 2026 average and recomputed in full at each publication. A model counts as launched only where every vendor pricing it that week was already tracked. Index behavior measures read each index’s own basket, so Repricing Incidence and Repricing Activity share one definition but are not expected to match.
Governance
Attic Standard administers every index it calculates, published or custom, under the practice expected of a price reporting agency.
PublicationPrices are dated to the Monday of their capture week and published once verified. Custom indexes reach their contract parties on the same schedule.
Queries and complaintsAny figure can be queried at info@atticstandard.com, or challenged with “Complaint” in the subject. Each is logged, acknowledged within five working days and answered in writing.
Changes to the methodA material change is announced with its reasons before it takes effect, and dated in the revision history once applied.
RecordsEvery price, decision, restatement, complaint and conflict declaration is kept for at least five years, so any figure can be reproduced.
Separate functionsPeople who sell or represent Attic Standard have no access to unpublished assessments and no role in capture, admission, calculation or publication.
Conflicts of interestA financial interest in, or paid role with, an indexed vendor or a buyer using the indexes is declared, recorded and excluded from every decision it touches.
Oversight committeeA committee with at least one independent member will be in place by March 31, 2027, followed by an independent review of IOSCO alignment.
Annual review and cessationEvery index is reviewed at least once a year. A discontinuation is announced six months ahead, or three where its market forces it, with a successor where possible.
Revision history
Date
Revision
History
August 2026
Weekly change became the geometric mean of individual price changes
Recomputed
September 14, 2026
One vote per model at both stages; May 2026 = 100; admission per model
Restated
September 18, 2026
Registry admission review; tier indexes; minimums from the base month
Restated
September 22, 2026
Market measures rebuilt on admitted prices; the nine redefined
Levels unchanged
October 5, 2026
Spot published as a series with its quartile band; base size disclosed
Levels unchanged
October 10, 2026
Index behavior measures defined; Repricing Incidence on lasting changes
Levels unchanged
October 27, 2026
Share of cuts in Repricing Activity now covers all lasting changes
Levels unchanged
A published figure is restated only where it is shown to be wrong and the correct figure is observable; otherwise it stands and is annotated in the record.
Scope and standards
A price benchmark earns its precision by measuring one thing. These sit outside it by design.
Outside the benchmark
Why
Performance and latency
Needs the models to be run, a separate discipline
Output quality
Specifications are recorded; models are not scored
Availability and rate limits
A listed price says nothing about capacity at that price
Total cost of ownership
Egress, storage, support and integration vary by deployment
What buyers actually pay
Negotiated rates are not public; list prices are their reference
Volumes and market share
No vendor publishes volumes, so nothing is weighted by them
Reading the indexes
Start datesA series starts once it meets its minimum constituents, and each card states its start.
Unconvertible unitsTracked but held outside the basket, so an index can hold fewer listings than its modality offers.
Retired modelsStay in modality indexes while resellers still sell them, until they are delisted.
Split billingWhere one request bills in several units, no single index holds its full cost.
Sticky list pricesA flat level beside a falling spot means the market moved through new models, not repricing.
Vendor-side errorsCorrected where the model’s history or a sibling listing shows it; otherwise captured as published.
IOSCO Principles
The indexes are administered in alignment with the IOSCO Principles for Financial Benchmarks (July 2013), the standard against which price reporting agencies in energy and commodities are assessed, applied in proportion to the administrator’s size and the risks its benchmarks carry. We describe this as alignment rather than compliance until the independent review has assessed it.
How Attic Standard meets each of the IOSCO Principles for Financial Benchmarks, applied in proportion to our size. Items in gold are commitments still being put in place.
Governance · Principles 1 to 5
1Overall responsibility of the administrator
Attic Standard is the administrator of every index it calculates, published or tailored to a client, and is responsible for every stage of each: capture, verification, calculation and publication.
§ 1.1, § 1.4 · In place
2Oversight of third parties
No third party contributes to a determination. Prices are captured by our own systems from vendors’ published sources.
§ 1.2 · In place
3Conflicts of interest
We hold no position, resell no capacity and earn no margin on any price we report. Commercial and assessment roles are separated, and conflicts are declared.
§ 1.1, § 1.5 · In place
4Control framework
Release only after complete capture, identity checks on every model and a continuous audit cycle.
§ 1.3, § 3.6 · In place
5Internal oversight
Every method decision is recorded with its reasons. An oversight committee with an independent member follows by March 31, 2027.
§ 1.4, § 1.5 · In progress
Quality of the benchmark · Principles 6 to 10
6Benchmark design
Each index answers one buyer question, built from medians and a chained matched-model calculation.
§ 2.1, § 3.3 · In place
7Data sufficiency
A census of published list prices, minimum constituents for every series and a two-observation rule.
§ 3.6 · In place
8Hierarchy of data inputs
Only list prices vendors publish themselves. No submissions, estimates or negotiated rates.
§ 1.1, § 2.3 · In place
9Transparency of determinations
Every value is calculated deterministically from recorded prices, and concentrated repricing is flagged.
§ 3.1, § 3.5 · In place
10Periodic review
Every published index is reviewed at least once a year, with a dated note.
§ 1.6 · New commitment
Quality of the methodology · Principles 11 to 15
11Content of the methodology
The full methodology is public on this page.
Parts One to Six · In place
12Changes to the methodology
Material changes are announced before they take effect and dated in the revision table.
§ 1.4, § 3.7 · In place
13Transition
Six months’ notice of cessation or material change, three where forced by the market, with a successor published in parallel where possible.
§ 1.6 · New commitment
14Submitter code of conduct
Our indexes use no submitted data.
§ 1.1 · Not applicable
15Internal controls over data collection
Deterministic extraction, a dry run before every write and verification before any price is recorded.
§ 1.2, § 1.3 · In place
Accountability · Principles 16 to 19
16Complaints procedures
Queries and formal complaints are logged, acknowledged within five working days and answered in writing.
§ 1.4 · In place
17Audits
An independent review of this alignment, once the oversight committee is in place.
§ 1.5 · Planned
18Audit trail
Prices, decisions, calculations, complaints and declarations are kept for at least five years.
§ 1.4 · In place
19Cooperation with regulatory authorities
Records are available to competent authorities on request.