A new real-market cryptocurrency trading experiment, which pits leading artificial intelligence models against each other to assess their investing skills, has so far seen a DeepSeek model outperform its competitors. Alpha Arena, launched last Friday by the US research firm Nof1, provided six large language models (LLMs) with US$10, 000 each to trade six cryptocurrency perpetual contracts on the decentralised exchange Hyperliquid, involving assets such as bitcoin and solana. By 2pm on Tuesday, DeepSeek’s V3. 1 led with a 10. 11 percent profit, while the lowest performer was OpenAI’s GPT-5, which suffered losses of 39. 73 percent. The other LLMs participating in this initial round of the experiment, which runs through November 3, 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. (Alibaba Cloud is the AI and cloud computing division of Alibaba Group Holding, which owns the Post. ) “Our goal with Alpha Arena is to create benchmarks that mirror real-world conditions, and markets are ideal for this, ” the Alpha Arena website stated.

They describe markets as “dynamic, adversarial, open-ended and endlessly unpredictable, ” adding that markets “challenge AI in ways that static benchmarks cannot. ”