Alpha Arena, a live cryptocurrency trading experiment organized by financial AI lab Nof1, is drawing attention for what it may reveal about the real-world capabilities of large language models in markets. The contest, which began on Oct. 18, places six AI chatbots in direct competition, each trading perpetual futures on Hyperliquid with an identical starting bankroll of $10,000.
The lineup includes Claude 4.5 Sonnet, Deepseek V3.1 Chat, Gemini 2.5 Pro, GPT-5, Grok 4, and Qwen 3 Max. Rather than evaluating these models on benchmark tests or paper portfolios, Alpha Arena focuses on a more demanding question: can AI systems make trading decisions with real money, manage risk, and maintain discipline under live market conditions?
Qwen and Deepseek Take the Early Lead
After eight days of trading, the leaderboard shows a sharp split in outcomes. Qwen is currently in first place, with its account rising to $17,496.35. According to the reported standings, it is holding just one active Bitcoin position. Deepseek sits behind Qwen and has also emerged as one of the strongest performers in the field.
By contrast, the remaining models have all lost money so far. The weakest result belongs to GPT-5, which has reportedly lost more than $7,000, placing it at the bottom of the competition. The divergence has fueled discussion among analysts and observers about the apparent performance gap between Chinese-built models and their Western counterparts in this specific trading environment.
A Live Test of AI Judgment and Risk Management
Alpha Arena is significant because it moves beyond theory. AI-driven trading is often described as a major future use case for machine intelligence, based on the idea that models can process vast amounts of information quickly and generate their own trading “alpha.” But live markets test much more than signal generation. They also expose weaknesses in consistency, position sizing, timing, and risk control.
That is exactly what Nof1 appears to be studying. The experiment is not only tracking profit and loss, but also how each model behaves when capital is on the line. This includes whether the systems reveal certain built-in biases, whether their trading styles differ in meaningful ways, and whether they can follow basic risk management rules over time instead of making isolated successful calls.
Jay Azhang, founder of Nof1, said the goals for this season include identifying the biases of each model, understanding the major differences in their trading styles, and evaluating whether they can adhere to foundational risk-management principles. In other words, the contest is as much about behavior and reliability as it is about returns.
Mixed Results Highlight the Limits of Current AI Trading
While the early standings may suggest an advantage for Asian models, the broader takeaway is more cautious. The experiment so far shows that automated AI trading remains an emerging field rather than a solved problem. Performance has been mixed, and the fact that only a subset of models is profitable underlines how difficult crypto derivatives trading remains, even for advanced language models.
The results also suggest that model quality in conversational or reasoning tasks does not automatically translate into superior market performance. Trading requires a blend of judgment, discipline, and adaptation to fast-changing conditions. A model that performs well in general-purpose tasks may still struggle in an environment where leverage, volatility, and execution decisions can quickly magnify mistakes.
This makes Alpha Arena useful as a practical benchmark. It offers a visible and ongoing comparison of how different AI systems respond to the same market structure, the same starting capital, and the same broad objective. For traders, developers, and investors watching the sector, that creates a rare window into where AI trading is making progress and where it still falls short.
Season Two Will Add a Human Trader
Nof1 has already teased a second season of the experiment. The next phase is expected to continue testing the featured models and may also involve the lab’s homegrown systems. One notable addition will be a human trader, giving observers a direct comparison between machine-led decision-making and traditional discretionary trading.
That comparison could become one of the most valuable parts of the project. If AI models can outperform not only one another but also a human market participant, the implications for algorithmic trading would be substantial. If they cannot, the limitations will be just as instructive.
Azhang struck a measured tone in his comments on the project’s future. He said the team is excited by the potential of large language models in trading, but remains skeptical and sees much more to test and learn. That caution matches the evidence so far: the promise is real, but the technology is still far from settled.
For now, Alpha Arena’s early data points to one clear narrative. In this live crypto trading contest, Chinese AI models are leading, Western rivals are lagging, and the broader race to build reliable AI traders is still in its experimental stage.