Editor's note: this article was first published in French on my Substack before being translated into English with the help of AI.

Can AI compete with humans in trading? Can its ability to analyze large datasets enable it to tame the volatility of financial markets?

This is what the Nof1 teams are working on with the Alpha Arena project. After cutting their teeth on cryptocurrencies, LLMs are now tackling stock trading.

For season 1.5 of Alpha Arena, US equities have become the battleground for 8 LLMs. The arena now welcomes Kimi 2, Moonshot AI's LLM, as well as a mysterious gladiator from a leading AI lab, alongside the main current models already competing in season 1.

In early November, Nof1 shook up the web3 ecosystem by launching the first cryptocurrency trading competition reserved for Claude 4.5 Sonnet, DeepSeek V3.1 Chat, Gemini 2.5 Pro, GPT 5, Grok 4, and Qwen 3 Max.

For this new competition, the rules haven't changed. Each model receives $10,000 of real money, in real markets, with the goal of maximizing trading profits over a 2-week period.

Each participant received more data than in the first season and benefits from identical input data to its competitors. The AI's scope of action has also been expanded as they compete in 4 distinct categories. In the New Baseline arena, LLMs have the ability to trade US equities based on news-related data as well as market sentiment. Monk Mode prioritizes capital preservation and strengthening risk management, while in Situation Awareness, models are aware that this is a competition, as well as their ranking and other models' performance. Finally, the Max Leverage arena pushes models to use maximum leverage on each trade.

With these 4 categories, Nof1 has therefore deployed $320,000 in capital for the needs of this new competition.

And as in the first season of Alpha Arena where Qwen 3 Max emerged as champion, only one model made profits across this entire new competition. This time it's the mystery model that achieved, across all categories, an average return on investment of 13.04%. GPT 5.1 comes in 2nd place by limiting its losses to -5.18%, followed by Gemini 3 which loses $2,842, or 28.41%. With $5,606 in losses, Grok 4 doesn't shine with its performance and finishes in last place in the competition.

Where does this mysterious model that managed to defeat LLMs already familiar with trading come from? Speculation suggests it could be the model developed by Nof1.

And for good reason, showcasing the performance of an AI in real conditions against industry giants is perhaps the best publicity the lab could hope for.

At this stage, it's difficult to compare AI's trading performance with that of humans. Not only because these experiments don't provide enough perspective on AI's ability to outperform the market over time, but also because it's difficult to estimate the average return of a human trader despite some studies on the subject.

Editor's note: this article was first published in French on my Substack before being translated into English with the help of AI.

Can AI compete with humans in trading? Can its ability to analyze large datasets enable it to tame the volatility of financial markets?

This is what the Nof1 teams are working on with the Alpha Arena project. After cutting their teeth on cryptocurrencies, LLMs are now tackling stock trading.

For season 1.5 of Alpha Arena, US equities have become the battleground for 8 LLMs. The arena now welcomes Kimi 2, Moonshot AI's LLM, as well as a mysterious gladiator from a leading AI lab, alongside the main current models already competing in season 1.

In early November, Nof1 shook up the web3 ecosystem by launching the first cryptocurrency trading competition reserved for Claude 4.5 Sonnet, DeepSeek V3.1 Chat, Gemini 2.5 Pro, GPT 5, Grok 4, and Qwen 3 Max.

For this new competition, the rules haven't changed. Each model receives $10,000 of real money, in real markets, with the goal of maximizing trading profits over a 2-week period.

Each participant received more data than in the first season and benefits from identical input data to its competitors. The AI's scope of action has also been expanded as they compete in 4 distinct categories. In the New Baseline arena, LLMs have the ability to trade US equities based on news-related data as well as market sentiment. Monk Mode prioritizes capital preservation and strengthening risk management, while in Situation Awareness, models are aware that this is a competition, as well as their ranking and other models' performance. Finally, the Max Leverage arena pushes models to use maximum leverage on each trade.

With these 4 categories, Nof1 has therefore deployed $320,000 in capital for the needs of this new competition.

And as in the first season of Alpha Arena where Qwen 3 Max emerged as champion, only one model made profits across this entire new competition. This time it's the mystery model that achieved, across all categories, an average return on investment of 13.04%. GPT 5.1 comes in 2nd place by limiting its losses to -5.18%, followed by Gemini 3 which loses $2,842, or 28.41%. With $5,606 in losses, Grok 4 doesn't shine with its performance and finishes in last place in the competition.

Where does this mysterious model that managed to defeat LLMs already familiar with trading come from? Speculation suggests it could be the model developed by Nof1.

And for good reason, showcasing the performance of an AI in real conditions against industry giants is perhaps the best publicity the lab could hope for.

At this stage, it's difficult to compare AI's trading performance with that of humans. Not only because these experiments don't provide enough perspective on AI's ability to outperform the market over time, but also because it's difficult to estimate the average return of a human trader despite some studies on the subject.