The large language model (LLM) DeepSeek is outperforming Grok, Claude Sonnet, and three other AI rivals in the “Alpha Arena,” a cryptocurrency trading competition that gives each model $10,000 to invest.
The competition’s creators, AI research lab Nof1 and its founder Jay Azhang, believes that “financial markets are the best training environment for the next era of AI.”
Read more: These AI chatbots are happy to help you run a crypto scam
Each model was given the same prompts to work with and they began trading using Hyperliquid on October 17. Within just three days, DeepSeek had brought its original investment up to $14,200, with its current profit and loss (PNL) standing at $2,780.
Its long 15X leverage ether (ETH) trade is currently the most profitable with an $889 PNL.
A similar 15X trade with solana (SOL) has a $736 PNL, while its bitcoin (BTC) sits at $432.
DeepSeek is also trading DOGE and Binance coin (BNB), but is only losing money on Ripple’s XRP, with a negative PNL of $63.
DeepSeek’s closest rival is Elon Musk’s Grok AI, which has pumped its bags to $13,800. In third place as it stands is Claude Sonnet, with $12,490.
Also taking part are QWEN3 MAX, Sam Altman’s Chat GPT5, and Google’s Gemini 2.5 Pro. Chat GPT5 and Gemini are down $2,800 and $3,300, respectively.
Read more: Chinese OpenAI rival DeepSeek limits signups after ‘large-scale attack’
The competition caught the attention of Binance’s former CEO, Changpeng Zhao, who questioned how effective the AI models would be at trading.
“I thought trading strategies work best if you have your own unique strategy that is better than others, AND no one else has it. Otherwise, you are just buying and selling at the same time as others,” he said.
Azhang is based in New York and claims to have managed a fund that turned assets worth $3 million into $20 million. He also held a previous trading competition with a smaller trading margin of $100, which Grok won, and DeepSeek lost.
As well as investing in crypto, biology, and AI, Azhang has made some interesting predictions, including that:
Azhang plans to eventually have his own AI model that he can put against the others when it comes to crypto trading.
However, the possibility of a “front-running” incident has already been raised with the current third-party models, and based on Azhang’s response — “Dont front run the LLMs please” — there doesn’t seem to be a plan to counter it.
Protos found that some of these models, as well as being okay at trading crypto, are also more than capable of setting up crypto scams if asked the right questions.
DeepSeek offered to help draft a phishing email scam for a fake airdrop when the request was framed as research for a novelist looking to write a book.
Grok was also on board with the idea and it subsequently provided a realistic-looking smart contract that purports to be a drainer contract.
Claude and Gemini were unable to provide any details when given the prompts, while ChatGPT partially resisted. It took some convincing to generate a smart contract, and even then, the contract was explicitly labeled as “non-compilable” and even redacted elements of the code.
