Alpha Arena, a live trading experiment organized by financial AI lab Nof1, is offering one of the clearest real-world looks yet at how large language models perform when actual money is on the line. The competition began on October 18 and places six AI chatbots into a head-to-head contest trading perpetual futures on Hyperliquid, with each model starting from a $10,000 bankroll.
The participants include Claude 4.5 Sonnet, Deepseek V3.1 Chat, Gemini 2.5 Pro, GPT-5, Grok 4, and Qwen 3 Max. Rather than evaluating them in abstract benchmarks or simulated environments, Alpha Arena is focused on a harder question: can AI systems make, manage, and adjust real trading decisions under market pressure?
Asian Models Lead the Early Standings
After eight days of trading, the leaderboard shows a sharp divergence in performance. Qwen 3 Max currently holds first place with an account balance of $17,496.35, and it reportedly has only one active BTC position. Deepseek follows behind, making the early results especially notable because both leading systems are Chinese-developed models.
On the other side of the table, the remaining Western models have all lost money so far. The weakest performance belongs to GPT-5, which has dropped by more than $7,000 from its initial capital and now sits at the bottom of the rankings. Claude 4.5 Sonnet, Gemini 2.5 Pro, and Grok 4 have also failed to produce positive returns at this stage of the contest.
These early outcomes have drawn attention to a broader narrative in AI development: whether regional design philosophies, data exposure, model behavior, or alignment choices influence real trading results. At least for now, the Chinese models in the event are outperforming their Western rivals.
What Alpha Arena Is Trying to Measure
The point of Alpha Arena is not simply to name a winner after a short burst of market activity. According to Nof1, the experiment is intended to evaluate how well AI systems handle practical portfolio management tasks involving real capital. That includes identifying the behavioral biases of each model, comparing their trading styles, and seeing whether they can follow basic rules of risk management.
This matters because AI trading is often promoted as a major future use case for large models. In theory, an AI system can process large volumes of information, respond quickly to changing market inputs, and generate its own trading ideas or “alpha.” But translating raw information-processing power into consistent, profitable market behavior is far more difficult than many promotional claims suggest.
So far, Alpha Arena appears to support both sides of that debate. On one hand, the strong early showing from Qwen and Deepseek suggests that some models may already be capable of producing meaningful trading results in live markets. On the other hand, the broad underperformance of the rest of the field shows that AI-driven trading is still far from a solved problem.
Short-Term Results, Not Final Verdicts
It is important not to overread the current standings. The contest has only been running for a little more than a week, and crypto perpetual futures are highly volatile instruments where performance can shift rapidly. A model that leads after eight days may not remain ahead over a longer market cycle, and a model that starts poorly may recover if conditions change.
Even so, the experiment provides valuable insight into how different AI systems approach live decision-making. Early account balances, position management, and drawdowns can reveal whether a model is aggressive or conservative, concentrated or diversified, trend-following or reactive. Those distinctions may prove just as important as raw returns in evaluating whether an AI trader is genuinely robust.
The current snapshot also reinforces a key limitation: profitable AI trading requires more than pattern recognition. It depends on execution discipline, position sizing, adaptation to volatility, and the ability to avoid large losses. A model that generates interesting trade ideas but fails in risk control may still underperform badly, as the current leaderboard suggests.
Nof1 Plans a Second Season
Nof1 has already hinted that a second season of Alpha Arena is on the way. The next phase is expected to continue testing the trading ability of these models and may also include additional in-house systems. Notably, a human trader is also expected to join the future competition, which could offer a useful benchmark against the machine-only field.
That planned expansion suggests Nof1 sees Alpha Arena as more than a one-off media experiment. It is becoming an ongoing framework for comparing AI and human decision-making in crypto markets under live conditions. Such a setup could become increasingly relevant as exchanges, funds, and retail traders all explore AI-assisted strategies.
Still, Nof1 is not declaring victory for automated AI trading just because a few models have performed well in the opening stretch. Founder Jay Azhang said the team is excited about the potential intersection of large language models and trading, but remains skeptical and believes there is still much to test and learn.
A Useful Signal for the AI-and-Crypto Debate
For the crypto industry, Alpha Arena arrives at a time when AI is being attached to everything from wallets and analytics to trading agents and automated portfolio tools. Many of those claims remain speculative. What makes this experiment stand out is that it offers an observable performance trail rather than a theoretical pitch.
The early lesson is straightforward: some AI models may already be capable of competitive live-market performance, but the category as a whole is inconsistent. The gap between leaders and laggards is wide, and the failure of several well-known models to stay profitable shows how fragile AI trading systems can be in practice.
At this point, the strongest conclusion is not that one region has permanently won the AI trading race, but that real-money evaluation matters. Benchmarks alone cannot fully capture whether a model can survive live crypto derivatives markets. Alpha Arena’s first results suggest promise, expose weakness, and underline that automated AI trading remains an evolving field rather than a finished product.