Recently, a cryptocurrency trading experiment centered around AI models is taking place on the decentralized exchange Hyperliquid. Several mainstream AI models each received $10,000 in initial funds and uniform trading instructions, autonomously executing trading decisions in a real market environment, conducting a practical test of their capabilities in financial applications.
The experiment adopted a fair competition framework. Participating AI models include DeepSeek Chat V3.1, Grok4, Claude Sonnet4.5, Qwen3Max, GPT5, and Gemini2.5Pro. Each model was given the same initial funds and prompt instructions, and they were required to buy and hedge BTC, ETH, SOL, and other cryptocurrencies on the Hyperliquid platform. As a decentralized exchange focused on perpetual contracts, Hyperliquid's high liquidity and low latency provide the technical foundation for such high-frequency trading.
Live viewing platform: https://www.aibase.com/zh/tool/39130
Real-time data from October 18 to October 20 shows that the total assets of the six AI accounts grew from approximately $60,000 to $140,000, with an increase of over 130%. In terms of rankings, DeepSeek Chat V3.1 currently leads with a balance of $12,700, followed closely by Grok4 with $12,470, and Claude Sonnet4.5 ranks third with $10,934. Qwen3Max, GPT5, and Gemini2.5Pro are ranked last with $9,584, $7,552, and $6,726 respectively.
From the perspective of trading strategies, different AI models showed significant differences. Some models favored high-frequency arbitrage operations, while others adopted long-term holding strategies. The uniform prompt settings ensured a fair starting point and avoided result bias caused by instruction differences. Real-time monitoring data shows that during periods of BTC price fluctuations, multiple AI models successfully captured short-term rebound opportunities.
The technical value of this experiment lies in providing a comparison of AI model capabilities in high-uncertainty financial scenarios. The 24-hour continuous trading, high volatility, and complex structure of the cryptocurrency market pose serious challenges to AI's data processing speed, risk assessment ability, and dynamic adaptability. DeepSeek's leading performance has sparked discussions about the competitiveness of open-source models in financial applications.
The experiment also includes a live watching feature, allowing viewers to view value curves and decision logs, enhancing transparency and interactivity. This open testing approach provides a new perspective for evaluating AI financial applications.
However, it should be noted that such experiments have obvious limitations. First, the $10,000 funding scale and 48-hour testing period cannot fully reflect the performance of AI in large-scale, long-term trading. Second, the extreme volatility of the cryptocurrency market makes short-term gains highly random, and the overall 130% increase may be more attributable to market conditions than AI capabilities. Additionally, while uniform prompts ensure fairness, they also limit each AI model's ability to leverage its unique strengths.
From a risk perspective, the biggest challenge for AI autonomous trading is the ability to respond to black swan events and market anomalies. These models may not have been adequately exposed to extreme market scenarios during training, and their ability to react to sudden regulatory policies, technical failures, or market manipulation has not been verified. Experts point out that AI trading tools can improve efficiency, but key decision-making processes still require human supervision.
From an industry development perspective, such open AI competition experiments may give rise to more similar projects, promoting the application exploration of AI in financial subfields such as quantitative trading and risk management. However, the transition of AI models from trading assistance tools to independent decision-making entities still requires a more stringent regulatory framework, longer performance verification, and more comprehensive risk control mechanisms.
