- Six large language models (LLMs) competed in the “Alpha Arena” crypto trading contest, with most finishing below their starting capital.
- ChatGPT lost 63% of its funds, leading the losses among competitors.
- Only two LLMs, DeepSeek and QWEN3 MAX, ended with profits, gaining $489 and $2,232 respectively.
- Trading costs significantly affected profits due to frequent trades and small gains.
- The competition setup imposed strict limits on models’ trading options and data access, making the task challenging.
Six large language models, including OpenAI’s ChatGPT, competed in the “Alpha Arena” crypto trading event created by Nof1. The contest, which ran for just over two weeks, involved all models trading cryptocurrencies under uniform prompts and conditions. The competition concluded on November 4th.
The final results showed that four of the six models ended with less than the initial $10,000 they started with. ChatGPT lost $6,267, equivalent to 63% of its funds. Google’s Gemini lost $5,671, X’s Grok lost $4,531, and Anthropic’s Claude Sonnet lost $3,081. Only High-Flyer’s DeepSeek and Alibaba’s QWEN3 MAX made profits, ending with gains of $489 and $2,232, respectively.
The number of trades varied widely among the models. Gemini executed 238 trades, while Claude Sonnet made only 38. Despite the difference in trade volume, all models had a win rate of 25 to 30 percent. Trading fees also impacted overall performance, with QWEN3 MAX paying the highest fees at $1,654 and Gemini paying $1,331 in transaction costs.
Nof1 highlighted that early losses were largely due to trading expenses. “PnL (profit and loss) was dominated by trading costs in early runs as agents over-traded and took quick, tiny gains that fees erased,” the organizer noted. On October 27th, some models briefly doubled their funds, and Claude and Grok managed short-term profits, while ChatGPT and Gemini remained mostly in the red throughout.
The contest was designed to be challenging, with LLMs receiving limited numerical time series data and restricted options for trading assets and actions. Nof1’s Jay Azhang commented that the models showed consistent behavior patterns, describing this as an investing “personality.” He added that the limited context and strict rules added difficulty.
Nof1 indicated a future competition is planned with improved prompts and added statistical rigor. Azhang’s goal is to develop his own crypto trading AI model, aiming to build on lessons from this event.
For further details, see Nof1’s complete post.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
Previous Articles:
- Appeals Panel Doubtful on Overturning SBF’s Conviction
- Critical OS Command Injection Flaw Found in React Native CLI
- Wall Street Raises AMD Price Target Ahead of Strong Q3 Report
- US Sanctions 50+ North Korean Crypto Addresses for Cybercrime
- Theta Labs Secures Patent for Advanced Tree-of-Thought AI System
