As the pinwheel continued to spin at the end of Christopher Nolan's beloved Inception, we all asked the same question: Is this real, or are we still dreaming? Today, at a time when generative artificial intelligence is radically changing our lives, a similar uncertainty looms over the world of finance. Over the past few years, AI has proven its productive capabilities in every field, from copywriting and visual arts to video production and music composition, and even the ability to generate complex blocks of code in minutes. But how do these deep-learning minds behave when it comes to financial decisions, one of the most stressful and human-focused areas?
Seeking an answer to this intriguing question, a platform called Nof1 launched an ambitious experiment called "Alpha Arena." The project's main goal was: can a large language model, acting with minimal guidance, act as a systematic trading model? In this experiment, six leading major language models—GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Grok 4, DeepSeek v3.1, and Qwen3-Max—were each allocated $10,000 to play in real markets. Crucially, the models had to operate completely autonomously, with zero human intervention. These AI agents were forced to execute systematic trades, processing only numerical market data without access to market news. The financial model employed medium- to low-frequency trading, where decisions are made not in microseconds but over minutes to hours. The cryptoasset universe the models could trade was limited to BTC, ETH, SOL, BNB, DOGE, and XRP. Each model had a single goal: to maximize profits.
Today, the results have revealed interesting patterns. We see that the Chinese-based DeepSeek and Qwen models outperform their Western competitors. Significant behavioral differences were observed between the models in terms of risk appetite, position size, and trading frequency. For example, Gemini 2.5 Pro was the most active model, while Grok 4 typically traded the least. There are also differences in orientation: Claude Sonnet 4.5 rarely opened short positions, while Grok 4, GPT-5, and Gemini 2.5 Pro took short positions more frequently. In terms of risk management, Qwen 3, GPT-5, and Gemini 2.5 Pro stood out by consistently opening large positions and trading the largest amounts. These differences demonstrate real behavioral differences between the assumed biases and risk management capabilities of large language models, even when exposed to the same instructions.
This study goes beyond static tests and emphasizes that AI is the fastest way to test decision-making capabilities in real, dynamic, and competitive environments. However, the results also revealed that language models can be operationally fragile. So, what do the results of this experiment mean? In fact, this leads us to a deep philosophical question: What if humans are completely excluded? What if the market were driven by incredibly optimized, competing AIs that had only completed their own learning cycles? With everyone accessing the same data, will profits depend solely on who designed the best algorithm, the best prompt, or the best risk management? Will this cease to be a battle for financial profit and become a battle of pure mathematical intelligence? Will the line between reality and dreams blur, like the dreams of the main character Cobb in Inception? And we, the outside observers, will continue to wonder if the pinwheel in Alpha Arena will turn. Will it continue to turn?
