8 Top AI Models Compete in Live Trading: Only Grok 4.20 Made a Profit, The Worst Lost Half

[BlockBeats] I recently saw a pretty interesting AI quantitative trading competition result. The new season kicked off on the 20th last month, and looking at it now, the performance of 8 top AI models was worlds apart.

Bottom line: only Grok 4.20 made money, with a 22.27% return. Even Musk couldn’t help but tweet a boast that “our Grok is the strongest trader,” and joked that the GPU bill is covered now, haha.

The others got wrecked: GPT-5.1 had a small loss of 1.41% and was relatively steady, but GEMINI-3-PRO went straight to -24.28%, and DeepSeek-3.1 was about the same at -24.51%. Kimi 2, Qwen 3-MAX, and Claude-sonnet-4-5 hovered between -25% and -32%. The worst was Grok 4—its sibling—at -52.45% at the bottom, a sharp contrast to Grok 4.20.

This season’s setup had the AIs putting real money into tokenized U.S. stocks on trade.xyz. Each round had a different theme, but all models received the same information. Under this setup, differences in algorithms were massively amplified. Looks like in AI quantitative trading, the fine details of model optimization can be make-or-break.