Chinese AI Models Lead in Crypto Trading Performance
Artificial intelligence models developed in China have outperformed US-based counterparts in cryptocurrency trading experiments, according to blockchain analytics firm CoinGlass. The findings come as top generative AI models compete across various domains, including the fast-moving crypto sector.
DeepSeek and Qwen3 Show Strong Performance
Data from CoinGlass highlighted that DeepSeek and Qwen3 Max, both created by Chinese teams, achieved the highest returns among several AI chatbots tested. DeepSeek stood out with a positive unrealized return of 9.1%, making it the only model in the experiment to post a net gain. Qwen3, developed by Alibaba Cloud, recorded a slight 0.5% unrealized loss but still ranked above many US competitors.
In contrast, Grok�a model with US origins�showed a 1.24% unrealized loss. OpenAI�s ChatGPT-5 performed the worst, with over a 66% loss, reducing its starting balance of $10,000 to $3,453 while the competition was underway.
Lower Development Costs, Better Returns
Observers were surprised by the success of DeepSeek, considering its training costs. DeepSeek was developed with a budget of just $5.3 million, based on the model�s technical paper. This contrasts sharply with the much higher expenditures for US models. For example, OpenAI reportedly spent $5.7 billion on research and development in the first half of 2025, and estimates for ChatGPT-5�s training range from $1.7 billion to $2.5 billion.
DeepSeek�s trading approach involved taking long positions on leading cryptocurrencies such as Bitcoin, Ether, Solana, BNB, Dogecoin, and XRP.
Possible Causes Behind Varying Results
Nicolai Sondergaard, research analyst at crypto intelligence firm Nansen, pointed to differences in AI training data as a probable explanation for performance gaps. He noted that while ChatGPT is a general-purpose model, others like Claude are tailored for coding. Strategic adviser Kasper Vandeloock added that prompt selection could impact results, especially for models like ChatGPT and Google�s Gemini.
- DeepSeek led with a 9.1% unrealized return.
- Qwen3 posted a marginal 0.5% loss.
- Grok delivered a 1.24% loss.
- OpenAI�s ChatGPT-5 faced a 66%+ loss.
Limitations of AI in Trading
The experiment began with $200 of starting capital for each chatbot, later raised to $10,000. Trades were executed automatically on the Hyperliquid exchange. While AI tools can detect market trends and signals, experts caution that traders should not rely solely on AI for autonomous trading decisions.
The latest results underscore both the rapid advancements in Chinese AI and the unpredictable challenges of applying AI to active crypto trading.
Related content
Comments
