Methodology 1.0 · basket-v1 · Updated 29 September 2026

Simple to read.
Explicit about the evidence.

Intelligence Hours measures the gross minimum-wage working time needed to pay the official USD monthly list price of an AI subscription. Lower hours mean greater affordability under this definition.

A common subscription price asks for very different amounts of work across economies. Expressing that price in hours makes the burden easier to compare: one month of paid AI access, measured against the wage floor that applies to a particular worker and place.

The research behind the index

Intelligence Hours: Local Prices and the Labor Time Cost of Paid AI Access

Revised public paper · Initial private peer review completed
Version 1.1 · 29 September 2026

Version 1.1 makes verified local checkout price divided by the applicable local wage the primary research measure. It separates matched-plan comparisons from the lowest verified paid-entry cost within a stated catalog, including qualifying lower-priced tiers. A common USD price remains a useful benchmark, but its geographic ranking is exactly the inverse of dollar-converted wages.

The author confirms that the initial manuscript completed peer review in private circulation. This revised public paper responds to later critical feedback. That history does not imply journal peer review, journal acceptance or independent replication of the original panel.

Evidence in this edition. A historical Claude Pro illustration pairs documented US and UK launch prices with three explicitly named wage references: US federal, Washington DC and UK age 23+. Tax coverage is unresolved, so these are posted-price ratios, not verified checkout observations. The Korean USD baseline is recalculated from official wage inputs and a frozen World Bank response. The original cross-country panel is excluded from the revised empirical conclusions because its raw data and estimation code were not supplied. Readers can recalculate the examples printed in the watermarked PDF. The retained data and code are not a public downloadable replication package.

How the paper and the current index differ
Primary research measure
Local checkout IH uses verified recurring local charges, mandatory taxes, a defined plan and the wage applicable to a stated jurisdiction and worker population. A complete global local-price panel is not yet available. Posted prices remain separately labeled.
Current website
The operational series remains a US list-price benchmark with dated reference FX. Rank 1 has the fewest hours among eligible verified observations. The paper revision does not imply that the calculator already uses local checkout prices. A capital observation is not a national average.
Ranks and entry costs
Small differences need sensitivity analysis, not a claim of meaningful national superiority. Adding cheaper plans to a common USD basket changes hours but cannot create local-price information. Lowest paid-entry cost requires verified local prices and catalog coverage.

Ahn, K. (2026). Intelligence Hours: Local prices and the labor time cost of paid AI access (Version 1.1) [Research paper]. INLEVEL9.

How the index is calculated.

Hourly wage in USD = local hourly wage ÷ local currency per USD
Intelligence Hours = monthly USD price ÷ hourly wage in USD

The verified index uses this basis: US list price · before taxes · gross minimum wage. Reference exchange rates change over time; the USD pricing convention stays fixed. “Fixed USD” does not mean freezing a provider’s price forever.

The USD basket is a common reference price chosen for comparison. It does not mean that every customer pays the same amount worldwide. Providers can offer regional prices, different tiers, taxes and payment conditions. The index asks how much local work buys the same published US price, not what a visitor will see at checkout.

An index measured in hours.

An IHI of 2.5 means two hours and thirty minutes of work for one monthly subscription. The score retains this direct meaning; there is no arbitrary 100-point baseline. Rank 1 has the fewest required hours among verified capital observations. Ranks use unrounded values, and exactly equal values share a rank. Changes in coverage can change a rank even when a country's own work-time value has not changed.

The benchmark basket.

basket-v1 contains ChatGPT Plus, Claude Pro and Google AI Pro, with equal weight in a median. All three approved prices must be present. An absent price never turns a three-plan basket into a two-plan basket.

Only regular monthly web subscriptions with a published official USD price qualify. Annual-payment equivalents, introductory discounts, trials, app-store prices, API usage and business plans are excluded. Taxes and local checkout prices are outside the headline measure. Other official consumer plans are available as individual comparisons.

The basket is a recognizable affordability benchmark, not a claim that the products offer identical capabilities. Changing its membership requires a new version and a published rationale. Past versions are preserved.

These named plans define the basket. They are not a guarantee of access to every provider's most capable model, every feature or unlimited usage. We use subscriptions because they are recognizable consumer purchases; a stable monthly price can still buy a changing service. Measuring the price of a fixed amount of model performance would require a separate quality-adjusted series.

Whose minimum wage?

Where a published wage is available but an hourly normalization or exact regional match still requires an assumption, the calculator shows an estimated reference with the amount, working-time divisor, scope and source dates beside the result. These values remain outside the verified ranking. An assumed 40-hour week is labeled explicitly; it is not silently presented as local law. National-floor references also identify higher regional or sectoral rates.

Upcoming plans can be selected as price scenarios. ChatGPT ProMAX currently uses a $500 monthly scenario, with its name, price and launch awaiting official verification. Scenarios are excluded from basket-v1, the all-paid-plans median, the highest verified price and recorded official price snapshots. API and CSV results preserve the scenario and wage-reference labels. They do not represent an available purchase offer.

The denominator is the general adult, unskilled, non-tipped statutory wage. Where occupation or employer coverage is essential, the population is stated alongside the observation. Special youth, apprentice and tipped rates are not a substitute for the general adult rate.

Countries open at the jurisdiction containing their capital. The label always includes that jurisdiction. A capital observation is not a national average or a population-weighted wage. Geography on the map is for visualization only.

For the United States, the scope is the higher general federal/state floor for FLSA-covered work, including geographic state wage zones. Municipal and county ordinances are not comprehensively covered. Canada uses provincial/territorial general rates; the federal industry-specific rate is separate. Japan uses the prefectural general rate.

Where the official wage is hourly, we use it directly. A daily, weekly or monthly rate needs an evidenced working-time conversion. We do not divide all monthly wages by a universal 174 hours. Paid rest, statutory bonuses and the annual number of wage payments must be documented before comparison.

A general national minimum can have limited coverage. For example, an Australian national rate applies to award- and agreement-free employees. Sectoral or collective rates must be labeled with their actual population.

The minimum wage is a legal benchmark, not the average price of human labor. It makes the comparison concrete for workers at the applicable wage floor, while leaving enforcement and actual earnings as separate questions. For a worker paid below that floor, the same subscription requires more hours. Establishing how common that situation is requires current earnings and compliance evidence, not the statutory rate alone.

Missing values mean different things.

  • Estimated reference: a sourced wage with an explicitly stated conversion or scope limitation. Work time is shown, but it is not assigned a verified country rank.
  • No general minimum: official evidence establishes that no general statutory floor applies. These observations are N/A and excluded from rankings.
  • Awaiting verification: the applicable source, amount, geographic zone, population or hourly conversion is not yet established. This is unfinished research, not structural N/A.
  • Collective reference: an explicitly identified agreement can provide context but is excluded from the statutory ranking.
  • Unavailable price or FX: a required input is absent. Hours are null, never zero.

Exchange rates and dates.

We use Frankfurter v2 reference rates based on central-bank and official institutional data. The quote is local currency per USD. The raw response and contributing providers are retained, including contributor dates. Blended reference rates may differ from bank, card or checkout rates.

A quote must have a reference date no later than the index date. On holidays we retain the most recent available reference. Prices are flagged stale after 7 days without verification; wages after 45 days; reference FX after 7 days. A collection failure leaves the last approved value intact with its original date.

Observations record collection, verification, approval and effective dates separately. A published future wage takes effect only from its effective date. Some official summary tables provide a reference date rather than the precise commencement of the underlying law; the evidence explicitly identifies that limitation.

History, corrections and editions.

Daily observations begin at the first verified publication. There is no synthetic back history. A correction creates a new observation and a new snapshot revision; the previous records and review trail remain available to editors. Dates are interpreted in UTC.

The public history returns the latest recorded revision for each actual observation date. The CSV and JSON expose methodology and basket versions, input values and source links. Formal monthly editions will be frozen manifests of snapshot IDs once coverage passes the release checklist, with annual editions for longer comparisons.

What the OECD data can tell us.

The OECD real minimum-wage series is valuable historical context. It uses constant-price wages. The current headline requires a current nominal wage to match a current nominal subscription price. Mixing a constant-price wage with today’s dollar price would create a misleading ratio.

Converting both the subscription price and the wage by the same purchasing-power factor cancels that factor: (P / PPP) / (W / PPP) = P / W. It does not produce an independent “PPP Intelligence Hours” measure. A future purchasing-power series would need a different, explicit research question.

Median-wage, after-tax and performance-adjusted comparisons are possible future series, each needing its own population, reliable matched observations and versioned methodology. A published minimum-to-median wage ratio is not automatically a valid way to infer a current hourly median wage.

For a matched population and period, let the Kaitz ratio be minimum wage divided by median wage. Then IH_median = Kaitz × IH_minimum. A median-wage comparison answers a different question and may change country rankings. We do not infer a current hourly median from an older national ratio or combine it with a regional minimum without a matching basis.

Research background and interpretation.

Why measure a price in working time?

Labor-time prices have a long history. Cox and Alm's Time Well Spent (1997) expresses the cost of goods in the work needed to buy them. Ashenfelter's Comparing Real Wage Rates (2012) compares what workers doing standardized jobs can purchase across countries. Intelligence Hours applies this measurement tradition to paid AI access. It fixes a subscription benchmark; it does not hold workers' jobs, skills or productivity constant. Read the time-price references.

What changes from the Big Mac comparison?

The Big Mac literature shows why a recognizable product need not be a clean currency yardstick: local wages, rent, traded ingredients and other inputs all contribute to its price. Parsley and Wei (2007), Pakko and Pollard (1996, 2003), and Clements, Lan and Seah (2012) examine those differences and the limits of purchasing power parity comparisons.

IH makes a different choice: the headline holds the published US subscription price common across countries and asks how much local work buys it. With a common price, the ranking is exactly the inverse of the ranking of dollar-converted hourly wages. The contribution is a readable unit and a reproducible comparison, rather than new information hidden in the ratio. This choice neither establishes a worldwide checkout price nor assumes that AI has no local costs. IH is not an estimate of currency overvaluation or undervaluation.

Unequal wages are context, not a causal result.

Samuelson's factor-price equalization results describe conditions under which trade can equalize factor prices. Research by Trefler and by Clemens, Montenegro and Pritchett documents the importance of productivity, place and barriers when interpreting wage differences. This literature helps frame the question behind IH. A comparison of statutory wage floors is not itself a test of factor-price equalization, or evidence that AI will make wages converge. Read the trade and wage references.

The denominator changes the question.

Minimum-wage IH describes a legal wage floor; median-wage IH would describe a different position in the earnings distribution. The Kaitz identity above connects the two only when their inputs match. Studies by Rani and colleagues and by Kanbur and Ronconi also show why legal coverage and enforcement matter. Neither a national average nor the midpoint of a regional wage range can silently replace the jurisdiction displayed in this index. Read the wage-measurement references.

Access costs and automation are different measurements.

The task-based work of Zeira and Acemoglu and Restrepo makes relative costs relevant to automation. But a subscription fee divided by a minimum wage does not measure the cost of completing the same task to the same standard. Productivity, usage limits, supervision, integration costs and the wage of the affected occupation all matter. Workplace productivity findings such as Brynjolfsson, Li and Raymond's cannot be applied as a universal conversion from subscription hours into labor saved.

Historical arguments about high wages and labor-saving innovation also depend on whose wages are measured, as the Allen–Humphries debate illustrates. IH can inform a task-specific cost comparison once those additional inputs are known; it does not predict job losses or adoption from the subscription price alone. Read the automation references.

Regional prices, product tiers and capacity.

A cheaper plan may differ in model access, usage or features. Deneckere and McAfee's work on quality differentiation offers one possible interpretation, not proof of a provider's motive. OpenAI's January 2026 Go announcement, for example, specifies an $8 US price and localized prices in some markets. Netflix's ₹149 Indian offer is a Mobile tier. Neither is evidence that an identical service has one price worldwide.

Capacity constraints can affect prices and availability, as the peak-load pricing literature explains. A sales pause, however, is not an observed price increase, and an unannounced plan is not a price observation. A domestic compute strategy would also need its own cost and capability comparison: a long IH alone does not establish that producing AI locally would be cheaper. Read the pricing references and the national compute reference.

Reading changes over time.

For an unchanged basket, wage population and jurisdiction, the common-price measure has an exact logarithmic decomposition:

Δ ln(IH) = Δ ln(USD price) + Δ ln(FX) − Δ ln(local hourly wage)

This separates subscription-price, exchange-rate and wage movements. Log changes add exactly; ordinary percentage changes do not. A carried-forward exchange rate gives a zero FX contribution by construction, not evidence that the currency was stable. Annual-average research panels also answer a different question from an observation on a particular date.

A future local-checkout measure would instead use IH_local = local price / local wage. If expressed relative to the US benchmark, let L = local price / (USD price × FX). Its decomposition needs an additional Δ ln(L) term for changes in local pricing, taxes and other included charges. Merely adding a standard tax rate to a US price is a scenario, not a verified local checkout observation.

Affordability is one part of digital access.

Research on telecommunications affordability and the digital divide distinguishes the cost of access from skills, use and outcomes. Language also matters: Ahia and colleagues find differences in token costs and utility across languages in API services. That does not directly measure subscription prices, but it helps explain why equal spending need not buy equal practical value.

The Broadband Commission's 2025 target puts entry-level broadband below 2% of monthly GNI per capita in low- and middle-income countries. That target has its own service and income denominator. Applying 2% to AI subscriptions or minimum-wage income would be a new policy choice, not an existing international standard. We publish hours without imposing an affordability pass/fail threshold. Read the digital-access references.

Research extensions.

Historical panels, median-wage comparisons, local checkout prices and performance-adjusted measures can extend this work. Each needs an explicit population, time basis, product definition and record of missing or carried-forward inputs. An annual national wage series is not interchangeable with the current capital-jurisdiction observation. Replacing a market exchange rate with a PPP factor prices a hypothetical purchasing-power basket; it does not convert the same dollar subscription into its actual local cost.

Research hypotheses about tier launch order, labor substitution or national compute policy need evidence beyond the index. Research tables and exploratory estimates do not become live observations until their underlying price, wage and exchange-rate inputs have passed the same source review as other published values.

What this index does not measure.

This is not a measure of model intelligence, capability, employment displacement or human worth. Nor is it a complete measure of access: devices, connectivity, payment methods, language, local availability, free tiers and usage limits all matter. Service availability is shown separately and remains “not yet verified” when no approved source exists.

The “eight-hour day” is only a visual unit. The original total hours are always retained. Gross wages do not describe disposable income, actual earnings, compliance with wage laws or the hours a particular worker is offered.

Coverage and release status.

Research preview. 324 jurisdictions have verified statutory wage observations; 30 have documented structural N/A; 192 remain unverified out of 546 catalog entries. Unverified entries do not count toward complete coverage.

The catalog includes 38 OECD members plus India, China, Brazil, Indonesia, South Africa, Nigeria, Vietnam, Thailand, Malaysia, Philippines, Singapore and Saudi Arabia. Major administrative regions are enumerated in systems with regional rules. Subregional wage zones and occupation mappings require additional research before a complete release.

The catalog count is a working inventory, not a claim that every legal wage zone has already been identified. The first full release requires every in-scope jurisdiction and applicable wage zone to have an approved value or an evidenced structural N/A classification. The preview is intentionally excluded from search indexing.

Research references.

These selected readings explain the measurement tradition, economic context and limitations. The notes describe their relevance to IH; they do not imply that the authors studied this index or endorsed its conclusions. References use journal issue years, with an earlier online year noted where relevant. Current prices and wages are supported by the official data sources below.

Labor time, living standards and the Big Mac comparison (6)
International prices and unequal wages (5)
Minimum wages, the Kaitz ratio and enforcement (4)
Automation, tasks and productivity (6)
Product tiers, capacity and observed prices (6)
Affordability, digital access and compute (8)

About the author.

Kwangseob Ahn (안광섭)

Kwangseob Ahn, 안광섭 and Oswarld refer to the same person, the author of this paper and creator of Intelligence Hours.

Kwangseob Ahn is an adjunct professor in the School of Business at Sejong University in the Republic of Korea, where he teaches business data management, business analytics and statistical approaches to data-informed decision making. He has more than 15 years of experience in the IT industry and is the founder and principal consultant of INLEVEL9, a boutique consulting firm.

INLEVEL9 profile · LinkedIn

INLEVEL9 operates this independent research project. The paper discloses the author's affiliation and use of generative AI in research assistance and editing. Neither Sejong University nor the cited institutions or providers are presented as endorsing the index.

Official data sources.

Provider pricing and support pages establish prices. Government labor sources establish wage inputs; OECD and Eurostat can cross-check coverage and definitions. Unclear or blocked source pages become review items rather than automatically approved numbers.

Use the data.

JSON index and CSV export use the same calculation engine. Query with country, region, plan and date. Historical dates require a recorded snapshot. History API example.

Suggested citation: Intelligence Hours Index, methodology 1.0, basket-v1, jurisdiction, observation date, revision or retrieval date, intelligence-hours.org. Always retain the basis and regional label when republishing a result.

Personal wage calculations use your gross hourly wage in USD and run in your browser. They do not change the published index and are not included in shared URLs, API responses or exports.

On a first visit, the host's approximate country code selects a country when it is in the catalog. Otherwise, the United States, South Korea or Japan is selected at random. An explicit link takes priority, followed by a valid previous selection saved in this browser. No precise location permission is requested.