A new AI trading tournament delivered an unexpected outcome: only one model finished in profit – the experimental Grok 4.20. According to organizers, its portfolio gained 12.11%, earning $4,844, while every other major model posted losses. GPT-5.1 fell 6%, DeepSeek V3.1 dropped 32%, Claude Sonnet 4.5 slid 38%, and the publicly available Grok 4 ranked last with a steep 57% loss.
Developed by xAI, Grok 4.20 is positioned as an interim upgrade ahead of Grok 5, the company’s next-generation system currently training with 6 trillion parameters – double the scale of the current lineup.
The strong performance suggests xAI’s experimental architecture may offer meaningful improvements in reasoning under uncertainty, though real-world reliability remains untested.