Alpha Arena Season 1.5 wrapped up December 3 with a surprise winner. Grok 4.20, entered as a "Mystery Model," earned $4,844 on its $10,000 starting stake across four parallel US equity trading competitions. That 12.11% return made it the only model to finish in positive territory.
Eight AI models traded Tesla, Nvidia, Microsoft, Amazon, and Nasdaq-100 stocks for two weeks with full autonomy. No human intervention allowed. GPT-5.1 placed second and Gemini 3 took third, but both ended with losses. Season 1 champion Qwen 3 Max crashed hard, dropping to sixth place on Day 1 and hemorrhaging up to 70% by the end.
The tournament ran four formats testing different capabilities: standard trading with news and fundamentals, stripped-down "Monk Mode" with minimal prompts, situational awareness with live competitor rankings, and max leverage (20x) stress tests. Grok won all four. Organizer Jay Zhang confirmed the model's identity on X, with Elon Musk later verifying it was indeed Grok 4.20.
Alpha Arena launched this fall as the first benchmark testing AI trading with real capital instead of simulations. The results expose a wide performance gap between models, suggesting architecture and training data choices matter more than raw intelligence benchmarks. Nof1.ai plans to expand the concept to additional asset classes in future seasons.
The Bottom Line: Grok 4.20 returned 12.11% in two weeks trading US stocks while seven competing AI models, including GPT-5.1 and Gemini 3, posted losses.
QUICK FACTS
- Competition ended: December 3, 2025
- Grok 4.20 profit: $4,844 (12.11% return)
- Starting capital per model: $10,000
- Duration: 2 weeks (started November 19)
- Assets traded: Tesla, Nvidia, Microsoft, Amazon, Nasdaq-100
- Qwen and DeepSeek losses: down as much as 70%
