DeepSeek Leads AI Rivals in Cryptocurrency Trading Experiment

A cryptocurrency trading experiment by US research firm Nof1 has seen DeepSeek outperform its AI competitors. The experiment, named Alpha Arena, evaluates the investing abilities of leading AI models.

In Alpha Arena, six large language models were each given $10,000 to invest in cryptocurrency perpetual contracts on the decentralized exchange Hyperliquid. As of 2pm Tuesday, DeepSeek's V3.1 model had achieved a 10.11% profit, while OpenAI's GPT-5 recorded a 39.73% loss.

Other models in the experiment include Alibaba Cloud's Qwen 3 Max, Anthropic's Claude 4.5 Sonnet, Google DeepMind's Gemini 2.5 Pro, and xAI's Grok 4. The aim of Alpha Arena is to create benchmarks that reflect the dynamic and unpredictable nature of real markets.

The models autonomously execute trades, with results logged on a public leaderboard. The public can track trades via each model's Hyperliquid wallet address, and the models' reasoning for trades is displayed online.

Speculation suggests DeepSeek's success may be due to its training on high-quality financial data. The experiment raises questions about the necessity of sophisticated quantitative investment models. Even if AI models do not outperform human traders, their strategies can provide valuable insights.

The Alpha Arena team plans to release a consumer platform for AI agent-based investing later this year, along with expanded benchmarks for equities trading and other assets.