News about AI often reads like a race between the United States and China. I wanted to see what the data actually shows. So we grouped the 136 models that have an independent score and a public price by the home country of the company that makes them.
The overview
| Country | Rated models | Best model (index) | Its rank | Median price | Open weights |
|---|---|---|---|---|---|
| United States | 79 | GPT-6 Astra (166.6) | #1 | $2 | 23% |
| China | 49 | Kimi K3 (157.7) | #12 | $0.68 | 80% |
| France | 7 | Mistral Medium 3.5 (141.4) | #91 | $0.75 | 57% |
| Canada | 1 | Command R (08-2024) (119.3) | #129 | $0.26 | 0% |
The rank is the model’s place among all 136 rated models on our overall index. The median price is per million tokens, for a typical mix of three tokens read for each token written.
The United States holds the top
All 10 best-rated models come from American companies, and the best of all, GPT-6 Astra, scores 166.6. American companies also have the most rated models: 79, from 8 companies. But they are the most expensive group: the median American model costs $2 per million tokens, and only 23% of them have open weights.
China competes on price and openness
China’s best model, Kimi K3, is number 12 in the ranking, 8.9 points behind the leader. The difference shows elsewhere: the median Chinese model costs $0.68, so the median American model costs 3× as much. And 80% of Chinese models have open weights: anyone can download them, run them on their own hardware or buy them from other hosts, which pushes prices down further.
Europe: one strong player
France is the only other country with a real presence: 7 rated models, all from Mistral AI. The best, Mistral Medium 3.5, scores 141.4 (number 91), and the median French model is the cheapest group in the table, at $0.75. Beyond that, Canada has a single rated model.
Does the country matter when you choose?
- Your data goes to whoever runs the model, not whoever made it. An open model made in China but run by an American or European host, or on your own computer, sends your requests to that host, never to its maker. If you use the maker’s own app or API, the maker’s terms and its country’s laws apply.
- Models carry the rules of their makers. Some Chinese models, for example, refuse or steer away from certain political topics. If your subjects are sensitive, test them first.
- For everyday work, compare scores and prices. The flag says little about how well a model will write your emails or read your documents. How much intelligence a dollar buys ranks the best value, whatever the country.
How we counted
The country is where the company that makes the model has its headquarters; Google DeepMind counts as American, as part of Alphabet. Only models with an independent score and a public price are compared here. Counting every model we list, rated or not, there are 162 American models and 126 Chinese ones. See all the makers on the companies page, and how open models compare in the open-weight gap.
The figures in this article are recalculated from our data at every update (last: Sep 30, 2026). How we work