nof1, an artificial intelligence (AI) research platform focused on financial markets, launched a large language learning model (LLM) trading test called Alpha Arena on October 18.
The test had six mainstream AI models (GPT-5, Gemini 2.5 Pro, Grok-4, Claude Sonnet 4.5, DeepSeek V3.1, and Qwen3 Max) use $10,000 in real funds each on the Hyperliquid crypto exchange, with identical prompts and input data. a
By the end of the experiment, DeepSeek and Grok had delivered returns of more than 14%, ranking in the top two. On the other end of the spectrum, Gemini 2.5 Pro had lost 42.57%.
Alpha Arena AI trading test
Unlike simulated backtests or paper trading, Alpha Arena ran entirely autonomously and in real time, measuring each model’s raw profit and loss (P&L).
All “contestants” were trading some of the most popular assets, including Bitcoin (BTC), Ethereum (ETH), and XRP. Uniform prompts ensured all models started from the same baseline, instruction-based bias.
The early leaders, DeepSeek and Grok, took aggressive long positions and capitalized on the ongoing market rally. In contrast, ChatGPT and Gemini, which mixed long and short positions, underperformed.
Overall, the Alpha Arena represents the first large-scale, public test of whether AI systems can genuinely interpret and react to live financial markets. A notable observation was that during Bitcoin’s sharp price swings, several models successfully identified and acted on short-term rebound opportunities.
Thus, the experiment offers valuable insights into how large language models handle high-uncertainty financial environments. However, it must be pointed out that a $10,000 portfolio and a 48-hour window cannot fully demonstrate long-term performance.
Similarly, the models were not really exposed to extreme market scenarios, leaving their crisis response ability untested. Still, the results have given developers a lot to think about in regard to how AI tools can improve trading efficiency and address human oversight issues.
Featured image via Shutterstock
