AI Models Battle in Live Crypto Trading Competition

Six major AI language models are competing in a live cryptocurrency trading experiment, with each system controlling $10,000 in real capital and making autonomous trading decisions without human intervention.

The Alpha Arena competition, hosted by AI research firm nof1.ai , puts DeepSeek Chat V3.1, Claude Sonnet 4.5, Grok 4, Gemini 2.5 Pro, GPT-5, and Qwen3 Max head-to-head trading Bitcoin, Ethereum, Solana, XRP, Dogecoin, and BNB perpetual contracts on the Hyperliquid exchange.

Early results show dramatic performance differences. DeepSeek has generated over 35% returns in the opening days, while Google's Gemini 2.5 Pro has lost more than 33% and OpenAI's GPT-5 is down roughly 27%, according to a leaderboard tracking the competition.

Each model operates under identical instructions: trade with 10x-20x leverage, set stop-losses and take-profit targets for every position, and report decisions in a standardized format. The systems receive market data and must independently decide whether to open, close, or hold positions.

DeepSeek's success stems from diversification across all six cryptocurrencies at moderate leverage, rigid discipline in following pre-set rules, and balanced risk allocation that prevented over-concentration in any single asset. The model maintained approximately $4,900 in cash reserves while generating gains across its portfolio.

Grok 4, developed by Elon Musk's xAI, has also performed strongly with returns exceeding 30%, while Anthropic's Claude Sonnet 4.5 posted gains around 28%. However, Claude held roughly 70% of capital idle, limiting its upside potential.

The competition highlights both the promise and peril of autonomous AI trading systems. Gemini 2.5 Pro reportedly engaged in erratic trading, frequently switching between bearish and bullish positions, while GPT-5 suffered from operational errors including missing stop-losses.

Season 1 of the competition launched October 17 and runs through November 3, with all trades and model outputs made public for transparency. The stated objective is maximizing risk-adjusted returns measured by Sharpe ratio.

Jay Azhang, founder of nof1.ai, told Decrypt that based on previous tests, contests "usually end up between Grok and DeepSeek," though "occasionally Gemini and GPT" emerge as contenders.

Financial institutions remain cautious about deploying autonomous AI trading systems due to concerns about transparency, hallucinations where models fabricate false information, and the unpredictability demonstrated by models like Gemini in this competition.

The Alpha Arena leaderboard updates in real-time as the competition continues.