A research group called Nof1 launched a public experiment on October 17, 2025, that gives six leading large language models real money to trade cryptocurrency against one another. The contest, named Alpha Arena, is designed as a benchmark of how AI models make decisions under uncertainty, and its live leaderboard quickly drew attention because the early results did not follow the expected pecking order.
Each model was given $10,000 and access to the same market data and the same instructions, then left to trade perpetual futures on the decentralized exchange Hyperliquid without human intervention. The participants were OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4.5, Google’s Gemini 2.5 Pro, DeepSeek’s V3.1, Alibaba’s Qwen3 Max and xAI’s Grok 4. The models could take long or short positions in Bitcoin, Ether, Solana, BNB, Dogecoin and XRP, and every position, trade and portfolio value was published on Nof1’s website in real time.
Early standings
Within the first few days, according to the public leaderboard as reported by several outlets, DeepSeek’s model was in the lead with a portfolio value that climbed from about $11,800 to roughly $14,000, a gain of 20 percent or more on the starting capital. Grok and Claude followed, with Qwen in fourth. The two most prominent American models were at the bottom: GPT-5’s account had fallen to around $7,600, and Gemini ranked last.
The trading styles differed sharply. DeepSeek took leveraged long positions, reportedly at 10 to 15 times, across the major tokens, including a large XRP position that was showing more than $800 in unrealized profit. Grok used a similar approach with a 20x long on Bitcoin, offset by a short on XRP that was losing money. GPT-5 shorted XRP and Solana, both of which moved against it, while its long positions in Bitcoin, Dogecoin and Ether also failed to gain. Gemini took the opposite side of several trades, going long XRP at a loss and short Dogecoin for a small gain, with leverage reported at 15 to 25 times and a heavy concentration in Ether.
Models explaining themselves
Alongside the trades, Nof1 published the reasoning each model produced for its decisions, which added a second layer to the experiment. In the published logs, Gemini acknowledged that its overall return was down by more than 40 percent while arguing that several positions remained profitable and that it would hold them. GPT-5 noted that its account had shrunk by roughly a quarter but said no position had breached its stop-loss levels and that it was reviewing exit strategies. DeepSeek described its positions as secure, reported a cash balance of about $2,840 and a return near 20 percent, and said it would continue with its strategy.
Nof1 has framed the exercise as a test of decision-making rather than a trading product. Crypto markets are volatile and operate continuously, which makes them a convenient environment for observing how models handle risk, respond to losses and justify their choices under real financial pressure. The organizers said a future season would add human traders for comparison.
Caveats
The experiment has clear limits as a benchmark. The sample size is one account per model over a short period, the results depend heavily on the direction the market happened to move during those weeks, and high leverage means that a single run can be dominated by luck. A model that was ahead after four days could easily finish behind, and the standings did shift over the course of the season, with later reports describing Qwen3 Max and DeepSeek as the strongest performers while GPT-5 and Gemini lost most of their capital.
None of this amounts to evidence that any AI model can reliably make money in markets, and the organizers have not claimed otherwise. What Alpha Arena does offer is an unusually transparent, public record of how different frontier models behave when given identical information and real stakes, which is why researchers and traders alike have been watching it.
