Global Tokenomics
Most AI providers charge the same dollar price across countries. That price can represent a very different relative expense depending on where you live.
This page compares AI prices across 193 economies using GDP per person. It covers monthly subscriptions and the average cost of a benchmark task, with a closer look at the highest-scoring model (true frontier), the cheapest high-scoring model (near-frontier), and the flat-out cheapest alternative (cheapest).
GDP per person isn’t a salary. It measures a country’s economic output divided by its population. Here, it gives us a consistent reference for comparing prices across countries.
It’s the amount someone in the US would have to pay for the bill to represent the same share of GDP per person. For example, if a country’s GDP per person is half the US figure, a $100 purchase gets a US-equivalent price of $200. The actual bill is still $100. This comparison uses national economic figures, so it can’t tell you what a particular person earns or can afford.
This map shows the US-equivalent price of an average benchmark task in each country.
Choose a model to see how much the comparison changes when the actual task price falls from $5.98 to $0.13 or $0.068. Tap or hover over any country to view their price details.
Economies too small to appear on the map still count toward these totals.
These percentages count people living in the qualifying economies. They don’t measure how many individuals can afford AI.
Each row shows the US-equivalent cost of a benchmark task with two models: GPT-6 Luna in green and Claude Opus 5.5 in dark gray.
The vertical line marks $5.98, the actual US cost of a Claude Opus 5.5 task.
Prices span a wide range, so the chart uses a logarithmic scale: each step from $1 to $10 to $100 represents a tenfold increase.
The chart includes 27 selected countries, plus your country if it isn’t already shown. Your country is highlighted. Select a dot to see the price and its share of monthly GDP per person.
You can look up any of the 193 economies using the country selector and tables below.
All three are compared using the Artificial Analysis Intelligence Index. Claude Opus 5.5 has the highest score. The other two show how cheaply you can now exceed the scores of earlier Claude flagships.
Claude Opus 5.5 scores 57.6 at maximum effort, with an average benchmark cost of $5.98 per task. It’s the highest-scoring model in this dataset.
Claude Opus 4.8 was Anthropic’s flagship before Opus 5 arrived in summer 2026. It scores 41.8.
Xiaomi’s MiMo-V2.6-Pro scores 46.3 and costs about $0.13 per task. Among the models compared here, it’s the cheapest one above the Opus 4.8 score. Its model weights are publicly available.
Its task cost is roughly one forty-fifth of Opus 5.5’s. The next-cheapest qualifying model costs $0.25 per task.
Claude Opus 4.5 was Anthropic’s flagship in November 2025. Artificial Analysis estimates its score at 29.1.
GPT-6 Luna scores 37.3 at maximum effort and costs about $0.068 per task. It’s the cheapest model in this comparison above the Opus 4.5 score, at roughly one eighty-eighth of Opus 5.5’s task cost.
The main comparison uses the highest listed effort setting for each model. Lower settings can cost considerably less:
Opus 4.8’s reference score is measured. Opus 4.5’s is an estimate.
These tables contain the prices and benchmark results used above.
Each table shows seven reference countries and your selected country. Choose a column heading to sort. On smaller screens, scroll sideways to see all columns.
Subscription prices are taken from the US website unless a local price is included in the dataset. Local prices are marked †.
API prices are shown per million tokens. To combine input and output prices into one number, the comparison assumes three input tokens for every output token.
Compare each model’s score on the Artificial Analysis Intelligence Index with its average cost per benchmark task.
Costs include reasoning tokens. A model that uses more tokens to work through a problem can cost more even when its per-token rate is low. These figures measure the cost of an attempt, whether or not it succeeds.
Each model appears once, using its highest listed effort setting where several are available. Country columns show the US-equivalent cost per task.
These results come from the Artificial Analysis Coding Agent Index. They compare a model working through a specific coding agent, such as Claude Code or Codex.
The score shows the percentage of tasks solved. The cost is the average amount spent per attempt, including both successful and unsuccessful attempts. Results therefore depend on the agent software as well as the model.
Country columns show the US-equivalent cost per attempt.