Alpha Arena, a live cryptocurrency trading experiment launched by financial AI lab Nof1, is beginning to reveal how leading large language models behave when real capital is on the line. The competition started on October 18 and places six AI chatbots in a head-to-head contest trading perpetual futures on Hyperliquid. Each model began with a $10,000 allocation, with the goal of testing not only profitability but also decision-making, portfolio construction, and adherence to risk management under real market conditions.
After eight days of trading, the early leaderboard shows a striking split between the top performers and the rest of the field. Qwen 3 Max currently holds first place with an account balance of $17,496.35, while maintaining only one active bitcoin position. Deepseek V3.1 Chat sits in second place. In contrast, the other participating models are all in negative territory so far, with GPT-5 posting the weakest performance and losing more than $7,000 from its starting capital.
Six Major Models, One Real-Money Test
The experiment brings together a notable lineup of frontier AI systems: Claude 4.5 Sonnet, Deepseek V3.1 Chat, Gemini 2.5 Pro, GPT-5, Grok 4, and Qwen 3 Max. Rather than evaluating these systems through benchmark prompts or simulated paper trading, Alpha Arena puts them into a live derivatives market where execution quality, conviction, and risk discipline directly affect outcomes.
That design is what makes the results especially interesting. In standard AI discussions, models are often judged on reasoning, coding, or text generation. Alpha Arena instead asks a different question: can these systems manage uncertainty well enough to trade real assets and preserve capital? The answer, at least from the first eight days, appears mixed. A small number of models have generated meaningful gains, but the broader field has struggled, suggesting that strong general AI performance does not automatically translate into trading success.
Chinese Models Have the Early Edge
One of the most talked-about outcomes so far is the regional divide in results. According to the standings cited in the experiment, the models developed in China have outperformed their Western peers in this early phase. Qwen and Deepseek are leading, while all of the Western models have lost money to date. That gap has drawn attention from analysts because it may reflect differences in model training, instruction tuning, response style, or how each system interprets market signals and balances aggression against caution.
At this stage, however, the findings should be treated as preliminary rather than definitive. Eight days is a very short period in crypto derivatives trading, especially in a market known for abrupt reversals and changing volatility regimes. A model that performs well over one stretch may falter in a different environment, and a lagging model could recover if conditions begin to favor its trading style. Still, the current standings are notable because they show that model-specific behavior can produce very different outcomes even when the participants are given similar resources and competing in the same venue.
Performance Is About More Than Returns
Nof1 has framed Alpha Arena as a broader experiment in AI competence, not simply a contest for the highest balance. The lab is using the event to study how different models express bias, how their trading styles diverge, and whether they can follow basic principles of risk management. Those goals matter because in real trading, profitability is only one layer of performance. A strategy that makes money while taking uncontrolled risk may not be robust, and a model that overtrades or ignores exposure limits can fail quickly even if some individual decisions are directionally correct.
The experiment therefore offers a practical lens on a fast-growing narrative in financial technology: that AI could become a major force in trading by processing large volumes of information and producing its own alpha. Supporters of AI trading often argue that advanced models can synthesize market data, news flow, and context faster than human participants. But Alpha Arena’s early results also underscore the limits of that thesis. So far, the field has not delivered uniformly strong outcomes, and the dispersion between winners and losers suggests that deploying AI in live markets remains a work in progress.
This is especially relevant in crypto, where perpetual futures are highly sensitive to leverage, sentiment swings, and liquidation dynamics. A model may identify an opportunity but still lose money if its sizing, timing, or risk controls are poor. In that sense, Alpha Arena is testing whether LLMs can move beyond analysis and operate as disciplined agents in a noisy, high-speed market environment.
Nof1 Remains Interested, but Cautious
Jay Azhang, founder of Nof1, said the current season was designed to identify each model’s bias, compare their trading styles, and determine whether they could follow basic portfolio and risk rules. That framing suggests the lab is less interested in a single flashy leaderboard snapshot and more focused on understanding where these systems are reliable, where they break down, and how they behave under pressure.
Azhang’s conclusion is optimistic but restrained. He said the team is excited about the potential intersection of LLMs and trading, yet remains skeptical because there is still much to test and learn. That caution aligns with the actual results: while the leaders have shown promise, the overall picture does not support the idea that AI trading is already solved. Instead, the experiment points to a future in which model selection, prompt design, execution constraints, and risk frameworks may be just as important as raw model intelligence.
Second Season Will Add a Human Trader
Looking ahead, Nof1 has already teased a second season of Alpha Arena. The next phase is expected to continue testing the trading skills of these models and will also include one of Nof1’s own homegrown systems. Most notably, a human trader is set to join the competition. That addition could make the project even more informative, because it would move the comparison beyond AI-versus-AI and toward a more practical benchmark: how machine-driven strategies compare with experienced human judgment in a live crypto market.
If that matchup takes place under similar conditions, it could provide one of the clearest public tests yet of whether LLM-based trading agents can truly compete with discretionary market participants. For now, the first season’s takeaway is straightforward: AI models can behave very differently in real-money trading, and early results suggest that some are adapting to crypto markets better than others. But the uneven performance across the field also makes one point hard to ignore—live trading remains a demanding environment, and today’s best-known AI systems are still learning how to navigate it.