Let AI models trade crypto with $10k each
Nof1's Alpha Arena gave six AI models $10,000 each to trade crypto on Hyperliquid with no human help; GNcrypto reports only Qwen 3 Max and DeepSeek finished in profit.
Done with Custom agentQwen 3 Max, DeepSeek, Claude Sonnet 4.5, Gemini 2.5 Pro, Grok, GPT-5
What they did
Nof1.ai ran Alpha Arena Season 1 from October 18 to November 3, 2025. Six models each got $10,000: Qwen 3 Max, DeepSeek v3.1, GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5 and Grok 4. They traded a type of crypto contract with no end date on the Hyperliquid exchange, with no human help. The organizers set the instructions and rules, and the models chose and placed their own trades.
How it went
GNcrypto reports Qwen 3 Max won with about 22% returns, finishing at $12,231. DeepSeek was the only other model to make a profit, ending at $10,489. GPT-5 did worst, ending with around $3,733.
Worth knowing
Nof1 founder Jay Azhang said the models only got streams of numbers over time, which these models don't handle well, plus few assets and limited actions. He also saw the same biases persist across many rewrites of the instructions.
Try it yourself
Act as a paper-trading agent for [asset list] using [data source]. Each [interval], review recent price data, decide whether to open, close or hold a position within a simulated balance of [starting amount], and log each decision with a short rationale. Do not place real trades or move any money. At the end of [test period], summarize performance, drawdowns and any recurring biases in your decisions. Show me your trading rules and logging format first and wait for my approval before starting.
Swap the highlighted [parts] for your own details.
Source: GNcrypto News
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
