AI Big Models Take on Crypto Trading Who Stands to Win

In the cryptocurrency market, human traders are often at the mercy of emotions and asymmetric information. But what happens when AI big models take the reins?

AI Models Go Live in Real Trading

On October 18, a project called Nof1 launched multiple AI models—including GPT-5, Claude Sonnet 4.5, Gemini 2.5 Pro, Deepseek V3.1, and Qwen3 Max—into live trading on the Hyperliquid platform. They traded major assets such as Bitcoin (BTC), Ethereum (ETH), Compound (COMP), Binance Coin (BNB), Dogecoin (DOGE), and Ripple (XRP).

Each model started with $10,000, aiming to maximize returns in the high-volatility crypto market. Unlike traditional simulations, Nof1 displays real-time price charts and account value curves, and includes a “BTC Holder” section for comparison against a simple buy-and-hold strategy.

By 11:00 AM on October 20, Deepseek V3.1 led with an account value of about $11,800. Grok, Elon Musk’s model, ranked second, Claude third, and Alibaba’s Qwen came in fourth. Surprisingly, OpenAI’s GPT-5 had an account value of only $7,600, second to last, while Google’s Gemini lagged at the bottom. These two are otherwise top players in the US App Store for large AI models.

Comparing Strategies

Deepseek employs an aggressive “max long” approach, leveraging 10-15x on all coins, currently showing unrealized gains across the board. Its large long position in XRP is unique among the models, yielding over $800 in floating profit.

Grok favors long positions as well but leverages BTC up to 20x and shorts XRP, the only position showing unrealized loss.

GPT-5 takes a more diversified approach, shorting XRP and COMP while holding long positions in BTC and ETH, yet most of its positions are currently underwater.

Gemini exhibits the most aggressive style, with leverage ranging 15-25x, heavy on ETH, and long XRP showing unrealized losses. Despite an overall 42% decline in portfolio value, Gemini maintains confidence, holding its positions firmly.

Deepseek noted that it still holds all six coins, with $2,840 in cash and a floating total return of 19.92%, letting predefined stop-loss and take-profit rules manage trades automatically.

What the Experiment Reveals

Nof1 is more than entertainment—it’s a stress test for AI in finance. The volatile crypto market provides a realistic decision-making environment, testing AI adaptability and robustness while driving algorithm refinement and commercialization of AI trading tools.

Key takeaways for investors:

1. Strategy drives returns: Aggressive leveraged positions can capture short-term gains but carry high risk. Diversified or hedged strategies are steadier but may underperform in the short term.

2. AI learns and adapts: Models adjust positions in real time based on market feedback, unlike traditional quantitative bots.

3. Volatility remains high: Investors should focus on risk management and understanding strategy logic, not just chasing leaderboard results.

My Perspective

Deepseek’s lead suggests that domestic quantitative teams in China have developed an edge in AI-driven crypto trading. Yet over the long term, market uncertainty and differences in strategy will continue to determine who truly wins. For investors, grasping each model’s risk appetite and strategy logic is far more valuable than blindly following short-term performance.

The next Nof1 season will introduce human traders and internally developed models, offering a further glimpse into the real competitive dynamics between AI and humans in crypto trading.