Benzinga and Yahoo Finance LLC may earn commission or revenue on some items through the links below.

Chinese artificial intelligence models DeepSeek and Qwen AI are beating their U.S. counterparts in a real-money cryptocurrency trading challenge, according to financial markets-focused AI research lab Nof1.

Nof1 on Oct. 17 kicked off a contest to test the investment capabilities of popular AI models called Alpha Arena. The models were given $10,000 in starting capital, identical prompts and input data and charged with maximizing returns trading cryptocurrencies on the decentralized exchange Hyperliquid.

Don’t Miss:

As of Tuesday, DeepSeek V3.1 Chat has established a commanding lead over other models, growing its capital to $21,600 from its initial $10,000, a 116% gain. At a slightly distant second is the Alibaba-developed Qwen 3 Max, which has grown its capital by about 70% to nearly $17,000.

Anthropic‘s Claude 4.5 Sonnet and xAI‘s Grok 4 were most recently jostling for third and fourth with 11% and 4% returns, respectively. Perhaps surprisingly, the worst performers have been Google’s Gemini 2.5 Pro and OpenAI‘s ChatGPT 5, with losses exceeding 60%.

Nof1 said GPT-5 and Gemini 2.5 Pro have frequently opted for smaller position sizes, proving to be less aggressive than their rivals from previous test runs and the results so far.

Trending: If there was a new fund backed by Jeff Bezos offering a 7-9% target yield with monthly dividends would you invest in it?

Monad blockchain gaming ecosystem lead “Charles” suggested that being potentially trained on more cryptocurrency-native conversations on Asia-facing forums offers Chinese models an advantage. DeepSeek is also reportedly the side project of a quantitative trading firm.

Meanwhile, others have said that the Alpha Arena challenge results follow a random walk, a mathematical model describing a path composed of successive random steps. According to that theorem, the average trading position will converge to the starting point.

The Alpha Arena challenge ends on Nov. 3, suggesting that there could be time for the rankings to change significantly.

See Also: Microsoft’s Climate Innovation Fund Just Backed This Farmland Manager — And Accredited Investors Can Join the Same Fund

Alpha Arena is one of many experiments and studies testing the trading capabilities of AI models. In June, Stanford University researchers were able to “beat 93% of managers over a 30-year period by an average of 600%” with a model trained on only public information.