A $20 AI Benchmark: 1 Work Hour in Luxembourg, 16 in Brazil

The Same $20 Takes Different Hours to Earn

ChatGPT Plus, Claude Pro, Google AI Pro. The paper uses these providers’ frontier-access plans as its reference and applies a common pre-tax monthly price of $20 to compare countries. That does not mean the actual checkout total is the same in Seoul, Berlin, and Lagos; local prices and taxes need separate observation. At that common $20 benchmark, the minimum-wage hours needed in 2025 were 1.1 in Luxembourg, 2.8 in Korea, 11 in Mexico, and 16 in Brazil.¹

I built a measure that expresses this gap in working time. I call it Intelligence Hours: the price of one month of AI subscription access divided by that country’s statutory minimum hourly wage, measured in hours. I calculated it using 2023–2025 data for the 37 countries whose minimum wages the OECD publishes, wrote it up as a paper (currently under review at the International Labour Review), and launched intelligence-hours.org, where you can compare countries directly.²

The trigger was September 10. OpenAI stopped accepting new sign-ups for its $200-a-month Pro plan, citing system load. It didn’t raise the price or lower it — it simply stopped selling.³ Here I also treat $200 as a common pre-tax comparison price. Earning that amount takes 12.7 hours in Germany, 27.6 hours in Korea, 98 hours in Mexico, and 752 hours in Nigeria.⁴ Even at one reference price, the labor time needed to earn it varies by more than tenfold.

Flip the Big Mac Index and You Get Intelligence Hours

The Economist has been building the Big Mac Index since 1986, and it was never really about the burger. It’s the same product sold in more than 100 countries, but its cost comes from local beef, local wages, and local rent — so comparing prices across countries reveals whether exchange rates are fairly valued. When researchers broke down the Big Mac’s cost, most of it turned out to be local labor.⁵

The comparison here reverses the approach. If the Big Mac Index measures “a local price for a global product,” the common-price version of Intelligence Hours measures “local labor spent buying a global reference price.” It puts the same $20 in each country’s numerator and compares the time needed to earn it. With that price held constant, differences come from dollar-converted minimum wages. The calculation does not establish that actual AI prices have converged worldwide or that local labor is absent from providers’ costs.

Of course, measuring prices in units of labor time is not new. A 2012 study already converted McDonald’s employees’ hourly wages into “Big Macs per hour” to compare real wages across countries.⁶ What drew my attention is the object being measured. Frontier models can assist labor or substitute for some tasks. The question here is how many hours of minimum-wage earnings are needed to pay for a month of access to those tools.

This is where automation and subscription affordability need to be kept separate. Assessing automation requires comparing machine and labor costs for equivalent output on the same task. A monthly subscription is not a task-specific cost, and the statutory minimum wage need not be the wage of the worker doing that task. Shorter Intelligence Hours means fewer hours of minimum-wage earnings are needed to pay for the specified subscription. It does not show that a machine performs the same work more cheaply than a person. Productivity, integration and supervision costs, and usage limits require separate evidence. The index does not measure actual subscriber numbers or adoption either.⁷

South Korea Is 0.5 Cents From the US Federal Minimum Wage

In the 2025 table covering 37 countries, the time needed to earn the common pre-tax $20 ranges from 1.14 hours in Luxembourg to 16.20 hours in Brazil — a 14-fold gap, with the median country sitting at 2.93 hours. Four distinct tiers emerge: nine countries across Western and Northern Europe plus Australia and New Zealand fall below 1.5 hours; the US federal rate, Japan, South Korea, and Poland sit at 1.5-3 hours; 11 countries across Central-Eastern Europe and the Mediterranean fall between 3-6 hours; and seven countries, mostly in Latin America, sit above 6 hours.⁸

Recalculating with the minimum wages taking effect on January 1, 2026, and adding China, India, and Nigeria from outside the OECD, the picture looks like this:

Country	Minimum Hourly Wage (USD)	Pre-tax $20 benchmark (hrs)	Pre-tax $200 benchmark (hrs)
Luxembourg	17.66	1.13	11.3
UK	16.08	1.24	12.4
Germany	15.71	1.27	12.7
France	13.58	1.47	14.7
Poland	8.35	2.40	24.0
Japan	6.84~8.19	2.44~2.93	24.4~29.3
US (Federal)	7.25	2.76	27.6
South Korea	7.26	2.76	27.6
China	2.43~3.85	5.19~8.22	51.9~82.2
Mexico	2.05	9.77	97.7
India	0.26~1.02	19.6~78.3	196~783
Nigeria	0.27	75.2	752

Exchange rates are 2025 annual averages. For Japan, China, and India, minimum wages vary by region, so I’ve listed the low-to-high range; the 2026 figures for India and Nigeria are unverified since I couldn’t confirm the original statutes. Adding Korea’s 10% VAT to the common $20 benchmark gives a scenario of 3.03 hours; this is not an observed local checkout total.⁹

Look at Korea’s row. The 2026 minimum wage of ₩10,320 (~$7.26) converts to $7.2551 per hour at 2025’s average exchange rate — a 0.07% difference from the US federal minimum wage of $7.25, less than half a cent. That’s a coincidence, and a currency-sensitive one at that: a 1% move in the won shifts the figure by 7 cents. Still, the picture is clear. Korea’s intelligence-hours are 2.16 times Germany’s.¹⁰ Korea is often called an AI powerhouse, but by the subscription-price burden relative to minimum wages, it sits in a different tier from Western Europe.

“Does the minimum wage really represent a country’s wage level?” is a fair objection. So I recalculated using median wages instead. Rankings shifted for 25 of the 33 countries, and the US dropped 10 places toward the cheaper end — because its federal minimum wage is pinned at just 24% of the median wage. Korea barely moved. Its minimum wage sits at 60% of the median, roughly in the middle, so switching to median wages does not greatly change Korea’s relative position. That does not identify the causal roles of policy and wages.¹¹

One sentence remains from all this. In this index, a higher subscription price and a lower wage both raise the same ratio. With the numerator fixed at $20, as it is here, cross-country differences come from dollar-converted minimum wages. Korea’s 2.76 hours is the time needed to earn that reference price, not a finding about its actual checkout price or productivity.

The FX market clawed back double what the minimum wage board gave

The real usefulness of an index lies in its time axis. Since the comparison holds the pre-tax numerator fixed at $20 for all three years, the change in Intelligence Hours splits precisely into two components — the exchange rate and the minimum wage. This isn’t an approximation; it’s an identity.¹²

Between 2023 and 2025, Intelligence Hours fell in 33 of 36 countries, because wage increases and exchange-rate changes together reduced the labor-time burden. Poland went from 4.11 hours to 2.79 hours; Croatia, Bulgaria, Romania, and Colombia all dropped by roughly a quarter to a third. Turkey’s currency lost over 40% of its value, but the country raised its minimum wage 2.2-fold, so its index actually shrank.¹³

Only two countries saw an increase: New Zealand, and South Korea. Korea raised its minimum wage twice during this period — from ₩9,620 (~$7.0) to ₩10,030 (~$7.3), a gain of 4.17 points on the log scale. Over the same span, the won weakened from 1,306 to 1,422 per dollar, a drop of 8.57 points. The net result: Intelligence Hours rose from 2.71 to 2.84 — but Korea’s rank fell from 14th to 18th out of 36 countries.¹⁴ For the burden of this common dollar-priced subscription, the foreign exchange market took back double what the minimum wage board handed out.

What follows is my interpretation and hypothesis, separate from the index’s empirical findings. Gold convertibility, dollar invoicing of oil, and AI subscription pricing are different institutions and transactions. My conjecture is that AI priced or ultimately settled in dollars could extend dollar dependence into cognitive work. Its magnitude, and who bears currency risk, depend on actual payment and settlement currencies, hedging, competing providers, and domestic alternatives. Local-currency checkout alone does not identify where that risk remains. I read America’s use of “Super Intelligence” as political language asserting technological leadership. Neither the label nor IH establishes an intention or an effect of strengthening dollar dominance.¹⁵

With the 2026 minimum wage set at ₩10,320 (~$7.5), the index would fall back to 2.76 hours — but only if the exchange rate holds at its 2025 average. The exchange rate that would erase this entire wage increase is 1,463.6 won per dollar.¹⁶ Minimum wages and exchange rates jointly affect the labor-time burden of this common price; in the two-year logarithmic decomposition, the exchange-rate term was twice the wage term.

The site I built shows the burden under the selected price and wage definitions each day. At the September 29 exchange rate, Seoul’s Intelligence Hours were 2 hours 38 minutes. That is shorter than the paper’s 2.76 hours, but comparing the site and paper requires checking their price and wage definitions as well as the exchange rate.¹⁷ To see how exchange rates affect subscription affordability, check this number again in a month while holding the price and wage assumptions constant.

Regional Pricing Wasn’t a Discount — It Was a Downgrade

Let me flip the question. Netflix, Spotify, and YouTube eventually moved to country-by-country pricing. Why did the U.S. reference price of AI subscriptions remain around $20 for over two years?

The explanation examined in the paper concerns the cost of additional use. Unlike a service where another stream costs little to deliver, frontier models consume GPU time and electricity as they generate tokens. That cost structure could constrain regional discounts on the same access tier. Offering a plan with reduced models or usage is one interpretation from the “damaged goods” literature. IH itself does not measure country-specific costs or providers’ pricing intentions.¹⁸

That’s exactly what happened. When OpenAI launched ChatGPT Go in India at ₹399 in August 2025, it wasn’t a price cut — it was a downgraded tier, with a lesser model and reduced usage limits. In the markets examined by the paper, regional discounts were observed on this reduced tier. If you calculate a “localization ratio” — local price divided by the US price for the same plan — Go comes in at 57–58% of the US price in India, Indonesia, and Nigeria, meaning a 43% discount. Plus, the frontier-access tier, was 115% of the US price in India, 106% in Indonesia, and 102% in South Korea. No frontier-tier discount was observed within that sample. South Korea stands out even further here: ChatGPT Go’s Korean price of ₩13,000 (~$9.3) is 114% of the $8 US price. Even excluding VAT, it’s 104%. Users in India pay 43% less for the downgraded tier; users in South Korea pay a premium for the exact same thing.¹⁹

The rollout order for Go fits this logic too. India (August), Indonesia (September), 16 Asian countries and 54 African countries (October) all came first, while the US, South Korea, Japan, and the UK came last, in the global launch of January 2026. The rank correlation between launch order and national income is 0.67. The exceptions are Norway, Denmark, Sweden, and Austria, which got it before the US — and that can’t be explained by income. Among rich countries, the order seems to be set by market size, payment infrastructure, and regulation.²⁰

So how much of the gap does localization actually close? On a log basis: 28% in India, 17% in Nigeria, 31% in Colombia, 27% in Mexico. That’s one-sixth to one-third. The rest is just the raw wage gap, unadjusted.²¹ These figures compare the price burden within the same Go plan; they do not measure its capability gap against Plus. Comparing Go and Plus prices in India also requires accounting for differences in the models and usage offered.

Oswarld’s Lens

The question I kept asking myself while building this index was “so what’s actually new here?” At a single point in time, common-price Intelligence Hours is simply the dollar-converted minimum wage turned upside down. I said as much in the paper. What I want to show is three things: a specified subscription access price expressed in minimum-wage hours; its exact exchange-rate and wage decomposition over time; and the cross-country variation in the numerator once local prices and taxes are observed. Automation thresholds and actual adoption require separate evidence.

I’ve always said not everyone, and not every task, needs superintelligence. There’s a model suited to each job and each setting. I myself hand off full planning and orchestration to Claude or GPT-6, but for long, repetitive work I’ve lately been offloading to DeepSeek or GLM. So I don’t think poorly of a stripped-down tier like ChatGPT Go on its own terms. The problem is that the gap doesn’t show up in the statistics. Counted by subscriber numbers, a Go subscriber in India and a Plus subscriber in Seoul both register as exactly one person. That is why subscriber counts need to be accompanied by the tier people can access. IH alone does not measure actual use or quality gaps, but holding the tier constant lets us compare the burden of its access price.

The debate over sovereign AI can also become more honest with this number. Beyond security or national pride, we can ask whether domestic provision can offer access at a specified capability level for a lower price. It is one possible route; competition, negotiated prices, efficiency gains, and subsidies also need comparison. Electricity, land, and people may be sourced domestically, while accelerators remain imports bought in dollars. The economics of domestic provision requires a separate comparison of delivered capability and full costs. Intelligence Hours identifies a larger subscription-price burden at the wage floor; it does not calculate the threshold at which domestic production becomes cheaper.²²

I’ll add one thing about Korea. At the same $20 reference price, a minimum-wage worker in Korea needs more working time to earn that amount than one in Western Europe. That is what this index shows. It does not identify who actually cannot subscribe or how effectively people use AI. The question I am answering here is: how many hours of minimum-wage earnings does a specified AI subscription cost? Keeping that boundary is what makes the number useful.

Closing

Economics has long measured wages by how many loaves of bread they can buy. This paper asks us also to look at how many working hours are needed to buy a month of AI access. In the 2025 comparison using a common pre-tax $20, that burden ranges from roughly one to sixteen hours, with Korea somewhere in between.

You can check the numbers yourself on the site. Pick a country, switch subscription tiers, plug in your own hourly wage. It recalculates daily using current exchange rates. The full methodology and sources are on the methodology page, and the paper’s data and code are all on Zenodo. If my calculations are wrong, feel free to catch me on it there.

---

💬 Reader, what AI subscription plan are you on right now? Have you ever worked out how many hours of your own wage that comes out to?

📨 If you know someone who, whenever the minimum wage comes up, always asks “so what can you actually buy with that money” — send them this piece.

References & Further Reading

• Ahn, K. (2026). Intelligence Hours: The Labor-Time Price of Frontier Artificial Intelligence in 37 Countries, 2023–2025. Public v4; under review at the International Labour Review. Cross-country figures follow the paper; site observations are dated separately.
• Ahn, K. (2026). Intelligence Hours: replication data and code for the 2023-2025 panel and exploratory 2026 extensions (1.0.0). Zenodo. The package that reproduces the paper’s panel, scripts, and tables.
• Intelligence Hours Index: a cross-country comparison site. Verified rankings for 25 countries as of September 29, 2026, updated with daily exchange rates.
• Methodology, sources and research: the site’s formula, basket definition (median of ChatGPT Plus, Claude Pro, Google AI Pro), and list of wage and exchange-rate sources.
• OpenAI Help Center: About ChatGPT Pro tiers: an explanation of the $100 and $200 Pro tiers.
• OpenAI sets ChatGPT Pro Max subscription at $500: coverage of the September 10 suspension of new $200 Pro sign-ups.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

• Intelligence Hours: The labor-time burden of a specified monthly AI subscription access price divided by the statutory minimum hourly wage. The cross-country comparison here uses a common pre-tax $20 benchmark based on the three providers’ frontier-access plans. Actual local checkout prices, taxes, and tiers require separate calculation.
• Kaitz ratio: The minimum wage divided by the median wage of full-time workers. It shows how high the minimum wage is set relative to a country’s wage distribution.
• Localization ratio: The locally listed price of the same plan divided by the exchange-rate-converted U.S. price. Below 1 means a local discount; above 1 means a local markup.
• Automation threshold: A comparison of machine and labor costs for equivalent output on the same task. IH does not measure this threshold.

---

각주

1. Table 1 of the paper. Based on 2025 OECD annual average minimum wages and World Bank annual average exchange rates. Luxembourg: 1.14 hours; Korea: 2.84 hours; Mexico: 11.04 hours; Brazil: 16.20 hours. ↩

2. Paper, Sections 3.1–3.3 (data and definitions). The site uses the minimum wage of the capital jurisdiction and same-day exchange rates, while the paper uses OECD annual averages and federal/national benchmarks, producing slightly different figures. ↩

3. OpenAI Help Center and September 2026 news coverage. Paper, Section 1. ↩

4. Paper, Section 1, Table 4. Based on 2026 minimum wages and 2025 average exchange rates. India ranges from 196 to 783 hours by state; Nigeria’s figure reflects an unconfirmed 2026 statute. ↩

5. Parsley & Wei (2007), cited in the paper’s Section 2.4. ↩

6. Ashenfelter (2012), cited in the paper’s Section 2.6. ↩

7. Public v4 of the paper, Sections 2.2 and 5.4. The task-based models of Zeira (1998) and Acemoglu & Restrepo (2018, 2019) motivate the research; they do not make IH a measurement of an automation threshold. ↩

8. Paper, Section 4.1, Table 1. ↩

9. Public v4, Sections 3.1 and 4.5, Table 4. $20 and $200 are common pre-tax comparison prices, not observed local checkout totals. ↩

10. Paper, Section 4.5. Germany’s 2026 Intelligence Hours figure is 1.27 hours. ↩

11. Paper, Section 4.2. U.S. Kaitz ratio: 0.24; Korea: 0.60. ↩

12. Paper, Section 3.5, Equation 3. ↩

13. Paper, Section 4.4, Appendix Table B1. For Türkiye, the exchange-rate term is +50.8 log points and the wage term is −79.8 log points. ↩

14. Paper, Section 4.4, Table 3. ↩

15. Public v4, Section 5.6. The connection between AI and dollar dependence is the author’s interpretation and research hypothesis, not an empirical finding of IH. ↩

16. Paper, Section 4.4. ↩

17. intelligence-hours.org, accessed September 29, 2026. Seoul: 2.63 hours, converted at a $7.6 hourly wage. ↩

18. Paper, Section 5.1. The “damaged goods” logic of Deneckere & McAfee (1996). ↩

19. Paper, Section 4.8, Table 6. Korea’s Go price is ₩13,000 (~$9.4), VAT included, as of April 2026. ↩

20. Paper, Section 4.7, Table 5. Spearman correlation of 0.67, based on 38 explicitly listed countries. ↩

21. Paper, Section 4.8. ↩

22. Public v4, Sections 5.3 and 5.5. Domestic production costs and capability require a separate comparison. ↩