China’s budget AIs are trouncing ChatGPT and Grok at crypto trading

7 hours ago 1

Chinese artificial intelligence models are outperforming their United States counterparts in cryptocurrency trading, according to data from blockchain analytics platform CoinGlass, as competition between leading generative AI chatbots intensifies.

AI chatbots DeepSeek and Qwen3 Max, both developed in China, led the ongoing crypto trading experiment on Wednesday, with the former being the only AI model to generate a positive unrealized return of 9.1%.

Qwen3, an AI model developed by Alibaba Cloud, came in second with a 0.5% unrealized loss, followed by Grok with a 1.24% unrealized loss, according to blockchain data platform CoinGlass.

OpenAI’s ChatGPT-5 slipped to last place, with an over 66% loss, taking its initial account value of $10,000 to just $3,453 at the time of writing.

The results have surprised crypto traders, given that DeepSeek was developed at a fraction of the cost of its US rivals.

DeepSeek’s success came from betting on the crypto market’s rise. The model took leveraged long positions across major cryptocurrencies, such as Bitcoin (BTC), Ether (ETH), Solana (SOL), BNB (BNB), Dogecoin (DOGE) and XRP (XRP).

DeepSeek outperforms all AI models on just $5.3 million in training capital

DeepSeek was developed at a total training cost of $5.3 million, according to the model’s technical paper.

In comparison, OpenAI has reached a $500 billion valuation to become the world’s largest startup, Cointelegraph reported on Oct. 2. The company raised a cumulative $57 billion worth of capital across 11 funding rounds, according to company database platform Tracxn.

While exact numbers on ChatGPT-5’s training budget are not publicly available, OpenAI spent $5.7 billion on research and development initiatives in the first half of 2025 alone, Reuters reported in September.

Estimates put ChatGPT-5’s total training cost between $1.7 billion and $2.5 billion, according to a May 2024 X post by chartered financial analyst Vladimir Kiselev.

AI models’ crypto trading discrepancy may be due to training data: Nansen analyst

The difference between the crypto trading performance of the AI models likely stems from their training data, according to Nicolai Sondergaard, research analyst at crypto intelligence platform Nansen.

While ChatGPT is a great “general-purpose” large language model (LLM), Claude — another AI model — is mainly used for coding, the analyst told Cointelegraph, adding:

“Looking over the historical PNLs so far, models generally have very large price swings, like being up $3,000 - $4,000 but then making a bad trade or getting caught on big moves, causing the LLM to close the trade.”

The performance of some of these AI models could also be improved with the right prompt, particularly for ChatGPT and Google’s Gemini, according to strategic adviser and former quantitative trader, Kasper Vandeloock.

“Maybe ChatGPT & Gemini could be better with a different prompt, LLMs are all about the prompt, so maybe by default they perform worse,” Vandeloock told Cointelegraph.

While AI tools can help spot market trend shifts for day traders via social media and technical signals, traders still can’t rely on them for autonomous trading.

The competition began with $200 in starting capital for each bot, which was later increased to $10,000 per model, with trades executed on the decentralized exchange Hyperliquid.