China’s low-budget artificial intelligence models are proving more efficient than their Western rivals in crypto trading, outperforming big-name systems like ChatGPT-5 and Grok, according to blockchain analytics platform CoinGlass. The trading experiment revealed that DeepSeek, a Chinese-developed AI model, was the only one to post a positive unrealised return on Wednesday, achieving a 9.1% gain despite its comparatively tiny development budget.
Following closely behind was Alibaba Cloud’s Qwen3 Max, which recorded a minor 0.5% unrealised loss. Elon Musk’s Grok fell further behind with a 1.24% loss, while OpenAI’s ChatGPT-5 performed the worst, seeing its account value fall from $10,000 to just $3,453—a loss of over 66%. The results stunned analysts, as DeepSeek was developed for only $5.3 million, a fraction of the billions invested in its American competitors.
DeepSeek’s trading strategy leaned heavily on bullish bets, taking leveraged long positions across major cryptocurrencies, including Bitcoin, Ether, Solana, BNB, Dogecoin, and XRP. This approach capitalised on the recent upward swing in the crypto market. In comparison, OpenAI’s ChatGPT-5—part of a company valued at around $500 billion—had vastly more resources but failed to turn a profit. Reports indicate that OpenAI spent $5.7 billion on research and development in the first half of 2025 alone, with ChatGPT-5’s training costs estimated between $1.7 billion and $2.5 billion.
Analysts suggest the gap in performance may come down to the type of data used to train each model. Nicolai Sondergaard, a research analyst at Nansen, noted that general-purpose language models like ChatGPT are not optimized for real-time trading decisions. He observed that such models often experience large portfolio swings—gaining thousands of dollars before losing them through poor timing or overexposure during volatile market moves.
Meanwhile, former quantitative trader and strategic adviser Kasper Vandeloock pointed out that results may also depend on how the models are prompted. “LLMs are all about the prompt,” he explained, adding that ChatGPT and Google’s Gemini could perform significantly better with optimised trading prompts or data feeds tailored for market analysis.
Despite the success of DeepSeek and the growing excitement surrounding AI-based trading, experts caution that these systems remain far from reliable for fully autonomous trading. While they can analyse social media trends, detect sentiment shifts, and interpret technical signals, their decision-making still lacks the contextual understanding and discipline required for consistent long-term performance. The experiment, which began with $200 per bot before scaling up to $10,000, was conducted on the decentralised exchange Hyperliquid.
