The first season of the Alpha Arena trading tournament, where AI models traded cryptocurrency for real money, ended in early November. Qwen3 Max from Alibaba won with a return of 22%, while GPT-5 from OpenAI performed the worst, losing more than half of its deposit. Now the organizers have summed up the results of the second season — this time the models traded stocks. The result is even worse: seven out of eight participants are in the red (important: after the end of the season, the models continue to trade on the exchange).
GPT-5.1, Gemini 3 Pro, Grok 4, Claude Sonnet 4.5, Qwen3 Max, DeepSeek V3.1, Kimi K2 Thinking, and a “secret model” from an unnamed laboratory. Each was given $10,000, and trading took place from November 20 to December 4 on the Hyperliquid platform. Instead of cryptocurrencies, tokenized shares of Tesla, Nvidia, Google, and Amazon were used. The organizers tested the models in four modes: basic, “monastic” (risk minimization), “situational awareness” (the model knows about the competition and sees the results of its competitors), and with maximum leverage of up to 20x.
The only model in the black was the secret one—its portfolio grew by 12.11%, bringing in $4,844 in profits. The organizers did not reveal who was behind it, referring to it only as “a model from a top AI lab.” It was displayed as a question mark on the leaderboard.
Among the public models, GPT-5.1 performed best, coming in second place but with a loss of 6%. Grok 4 (-57%) and Claude Sonnet 4.5 (-38%) performed the worst. Qwen3 Max, the winner of the first season, lost 29%. DeepSeek V3.1 lost 32%. An interesting detail: DeepSeek showed +16% in “monastic” mode, but failed in the others — it seems that risk management settings are more important than the model itself.
The main question remains open: who created the secret model that beat everyone? The organizers promised to reveal the developer after the tournament ended, but so far they remain silent.