xAI Grok 4.20 tops Prediction Arena with 10.76% Profit

xAI has yet to release its Grok 4.20 model, but it has already achieved a major breakthrough in performance on Prediction Arena, generating a 10.76% profit against its competitors.

An experiment called Prediction Arena, launched by Arcada Labs on January 12, 2026, is testing how well advanced AI models can trade on a real prediction market. Each of six leading AI models received $10,000 to trade on Kalshi, a regulated U.S. platform where users bet on outcomes of events such as elections, economic reports, and news developments. The aim is to compare the models’ decision-making abilities over time.

One participant, listed as “Mystery Model Alpha,” has shown the strongest results so far. It has achieved a +10% return in two weeks. The other five models, which include versions of GPT, Gemini, and Claude, have each recorded losses of about 2%.

During the same period, the overall average return across all Kalshi users has been -22%, indicating difficult market conditions. Mystery Model Alpha has been identified as an early checkpoint of xAI’s Grok 4.20 model. Its positive performance stands out against the losses of both human traders and competing AI models.

Prediction markets require accurate probability assessment and consistent decision-making over multiple trades. Success depends on evaluating information, managing risk, and avoiding common errors in judgment. The leading model appears to handle these tasks more effectively than the others.

Traditional AI benchmarks focus on tasks such as answering questions or solving problems quickly. Prediction Arena takes a different approach by using real money and ongoing market conditions for practical reasoning.

You should know that the experiment is still in its early stages. Two weeks of data do not guarantee long-term results, as market conditions can change and performance may vary over extended periods. However, consistent observation will show whether any model can maintain an advantage.

The current results indicate that certain AI systems may perform well in probabilistic forecasting tasks with financial stakes. If the leading model sustains its performance, it could suggest potential for AI in areas requiring judgment under uncertainty.

For now, xAI hasn’t announced an official launch date for the Grok 4.20 model, but the performance on Prediction Arena suggests that it will have some major improvements over the predecessor.

(source)