In a surprising turn of events, two Chinese AI models claimed the top spots in the “Alpha Arena” investment competition, while major US models suffered significant losses, with GPT-5 ranking last after losing over 62% of its capital.
Real Money, Real Trading
Organized by startup Nof1, this wasn’t a simulated exercise. To genuinely test AI investment capabilities, each model received $10,000 in real startup capital to autonomously trade digital currencies in live markets. Alpha Arena broadcast the entire process—showing real-time price fluctuations, live profit rankings, and each model’s trading rationale.
The Final Standings
After 17 days of intense competition, here are the results:
-
Alibaba’s Qwen3 Max – 22.32% return ($12,232 balance)
-
DeepSeek’s chat v3.1 – 4.89% return ($10,489 balance)
-
Claude Sonnet 4.5 – Over 30% loss
-
Grok 4 – 45% loss
-
Gemini 2.5 Pro – Over 30% loss
-
GPT-5 – 62.66% loss ($3,734 balance)
Notably, the Chinese models were the only participants that managed to turn a profit.
Trading Styles Revealed
Each AI displayed distinct trading personalities:
-
DeepSeek chat v3.1 demonstrated remarkable stability, leading for most of the competition. Its “rational” approach involved broad diversification with no position changes—no stop-losses, no profit-taking. This aligns with its origins; DeepSeek‘s parent company,幻方 (Huan Fang), is a quantitative trading firm.
-
Qwen3 Max employed a surprisingly simple strategy: daily “all-in” bets on single assets with multiple leverage. While this approach caused significant losses during directional mistakes, it ultimately delivered the highest returns.
-
Grok 4 showed aggressive, trend-chasing behavior with high-frequency adjustments, resulting in substantial volatility.
-
Claude excelled at analysis but struggled with execution, frequently hesitating and failing position adjustments.
-
Gemini 2.5 was humorously compared to retail traders—constantly changing strategies (long/short flipping) and generating high transaction costs.
The Bigger Picture
Nof1’s vision extends beyond this competition. They believe financial markets represent the next frontier for AI development—a training environment that becomes increasingly challenging as AI grows more sophisticated.
“We use markets to train new foundation models,” the team stated, aiming for AI evolution through open-ended learning and large-scale reinforcement learning to tackle ultimate complex challenges.
But Is AI Trading Really Reliable?
Financial professionals remain cautious. AI lacks understanding of individual investors’ actual financial situations, family circumstances, employment status, and risk preferences. Providing investment advice without this context could be risky. Furthermore, AI’s fundamental logic revolves around summarizing and reproducing existing human knowledge—it doesn’t predict the future.
The ideal approach might combine rational AI tools with human wisdom, creating a powerful partnership rather than complete automation.
*Keywords: AI investment competition, Alpha Arena, DeepSeek trading, Qwen3 Max performance, GPT-5 losses, AI trading styles, autonomous trading AI*