Chinese AI models are surpassing their US counterparts in cryptocurrency trading, as reported by blockchain analytics platform CoinGlass, amid escalating competition between top generative AI chatbots.
On Wednesday, the ongoing crypto trading experiment saw AI chatbots DeepSeek and Qwen3 Max from China leading the way, with DeepSeek achieving a notable positive unrealized return of 9.1%.
Following closely, Alibaba Cloud’s Qwen3 recorded a 0.5% unrealized loss, while Grok managed a 1.24% unrealized loss, according to CoinGlass data.
In stark contrast, OpenAI’s ChatGPT-5 fell to the last position, incurring an over 66% loss, reducing its initial account value of $10,000 to merely $3,453 at the time of writing.
The results have taken crypto traders by surprise, particularly since DeepSeek was developed at a fraction of the cost of its US competitors.
DeepSeek’s success is attributed to its strategy of betting on a rising crypto market. The model executed leveraged long positions in major cryptocurrencies, including Bitcoin (BTC), Ether (ETH), Solana (SOL), BNB (BNB), Dogecoin (DOGE), and XRP (XRP).
Related: Arthur Hayes calls for $1M Bitcoin as new Japan PM orders economic stimulus
DeepSeek outperforms all AI models on just $5.3 million in training capital
DeepSeek’s total training cost stood at $5.3 million, according to the technical paper of the model.
In comparison, OpenAI has achieved a valuation of $500 billion, becoming the world’s largest startup, as Cointelegraph reported on October 2. The company has raised a total of $57 billion through 11 funding rounds, according to Tracxn.
Though exact figures on ChatGPT-5’s training budget are undisclosed, OpenAI reportedly spent $5.7 billion on initiatives in the first half of 2025 alone, as noted by Reuters in September.
Estimates of ChatGPT-5’s total training costs range between $1.7 billion and $2.5 billion, according to a May 2024 post by chartered financial analyst Vladimir Kiselev.
Related: $19B market crash paves way for Bitcoin’s rise to $200K: Standard Chartered
AI models’ crypto trading discrepancy may be due to training data: Nansen analyst
Nicolai Sondergaard, a research analyst at Nansen, suggests that the differing performances of the AI models in crypto trading likely stem from their respective training data.
While ChatGPT serves as a robust “general-purpose” large language model (LLM), Claude focuses primarily on coding tasks, the analyst informed Cointelegraph, adding:
“Reviewing historical PNLs thus far, models exhibit significant price fluctuations, often experiencing gains of $3,000 – $4,000 followed by poor trades or adverse market movements that lead LLMs to close trades.”
Moreover, the efficiency of some AI models could be enhanced with proper prompts, especially for ChatGPT and Google’s Gemini, as stated by former quantitative trader and strategic adviser, Kasper Vandeloock.
“Perhaps ChatGPT & Gemini could perform better with different prompts; LLMs greatly depend on the prompt, potentially leading to suboptimal default performances,” Vandeloock mentioned to Cointelegraph.
Although AI tools can assist day traders in identifying market trend shifts through social media and technical signals, they are not yet reliable for autonomous trading.
The experiment commenced with each bot starting with $200, which was subsequently raised to $10,000 per model, with trades carried out on the decentralized exchange Hyperliquid.
Magazine: Crypto traders ‘fool themselves’ with price predictions — Peter Brandt
