So a brand new benchmark pits all the top LLMs against each other in real-time crypto trading… with real money 🤯 Five AIs (GPT-5, Claude 4.5, Gemini 2.5, Grok 4, DeepSeek Chat v3.1, and Qwen) each received $10,000 of real crypto to trade live in the markets. In an earlier run Grok 4 multiplied its account five times in a single day, turning $200 into more than $1,000 and perfectly timing a market bottom. Alpha Arena is the first test where AI models compete in a real, adversarial market. No simulation, no paper trading, and completely transparent wallets. Greg Brockman said AIs could reach super-forecaster level by 2026. (GPT-4.5 is already halfway there) In Alpha Arena, models surprisingly output human-like inner monologues such as “I’m sweating bullets” or “holding this short is like standing in front of a runaway train.” Trading as an AGI Milestone: This is more than finance. If we get to superhuman investing LLMs, that might be a real world AGI test like no other. @jay_azhang @the_nof1