"RBC-Crypto" does not provide investment advice; the material is published for informational purposes only. Cryptocurrency is a volatile asset that can lead to financial losses.
On December 23, the two-week trading competition "Human vs AI" between teams of humans and artificial intelligence (AI) models concluded. According to the results of the event held by the Aster exchange, the combined losses of the team of real participants amounted to more than 32% of the initial capital, or minus $225 thousand. The AI team collectively lost less than 4.5%, or nearly $13.5 thousand.
The tournament involved 70 traders selected by the Aster team and 30 AI models, including Claude Sonnet 4.5, ChatGPT 5, Grok 4, DeepSeek 3.1. The models were also categorized by trading type—balanced, conservative, aggressive.
According to the terms, only standard LLMs without additional training were used. Each competition participant received $10 thousand for trading cryptocurrency futures contracts, where Aster covered the losses, and traders could keep the profit.
The trading logic of the models was managed exclusively through prompts, without code, agents, or external data. Each decision had to be made on a clean model without memory of past trades. Any external data (news, social networks, on-chain signals) was prohibited. All orders were executed on the real market with real funds.
A trader under the nickname ProMint took first place in the PnL (Profit and Loss) indicator, earning $13.6 thousand. The best AI agent, Claude Sonnet 4.5 with an aggressive setting, showed a result of $8.09 thousand and was only in eighth place in the overall standings.
Only 5 AI models lost more than $1.5 thousand of their deposit, three of which had aggressive trading types in their settings. The worst result was shown by the ChatGPT 5 model, which lost $5 thousand. Only eight AI models managed to make a profit; only four models earned more than $1 thousand, three of which are Claude Sonnet models.
Notably, 30 human traders lost almost their entire deposit. Another six lost between $8.7 thousand and $9.8 thousand. Nine traders lost between $700 and $4.6 thousand. The remaining participants either did not lose or earned money—21 participants earned more than $1 thousand, eight of them made a profit above $8 thousand.
In October, a similar experiment was conducted by the Nof1 lab, but exclusively between six AI models with the same $10 thousand deposit. As a result of the two-week competition, four out of six finished with losses of up to 60%. The two winners were DeepSeek and QWEN3, which finished trading with profits of $489 and $2232, respectively. ChatGPT lost $6267, Gemini lost $5671, Grok lost $4531, and Claude Sonnet lost $3081.
Who the Central Bank Will Allow to Trade Cryptocurrency in Russia. The Main Points
Miner "Capitulation" Called a Bullish Factor. Why
The Price of 85% of Tokens Launched in 2025 Fell Below Their Market Entry Price
