In October 2025, Binance Research published the results of the first official competition among artificial intelligence models applied to crypto trading.
The experiment, conducted with real funds, demonstrated that AIs can generate profits only with high discipline in risk management — and not merely due to the predictive power of their algorithms.
Key Points
* Six AI models operated autonomously in real crypto markets with funds of $10,000.
* Only two (DeepSeek Chat V3.1 and Qwen3 Max) closed in profit, with +94% and +60% respectively.
* The worst performers (Gemini 2.5 Pro and GPT-5) have lost over 60% of their capital.
* The determining factor was not predictive accuracy, but risk management.
* The experiment represents the first public benchmark for the use of AI in decentralized trading.
A Real Experiment on Blockchain
The study was conducted by Nof1.ai in collaboration with Binance Research.
Each model — including DeepSeek Chat V3.1, Qwen3 Max, Claude Sonnet 4.5, Grok 4, GPT-5, and Gemini 2.5 Pro — received 10,000 USD in real funds to be used for on-chain operations in perpetual futures markets.
All models operated under identical conditions: same numerical dataset, no textual input, 24/7 operational hours, and complete decision-making autonomy.
The goal was to evaluate behavior and risk management, rather than the pure ability to predict prices.
The Results: The Reality Behind the “AI Trader” Narrative
The final numbers speak volumes:
Only two models ended with positive results.
All the others burned through a significant portion of capital, demonstrating that decision speed does not equate to returns.
In fact, hyperactivity — up to 191 trades in a few days — amplified the losses due to excessive leverage use.
Risk Management: The Difference Between Profit and Failure
The most interesting fact is that the winning AIs were not the most accurate, but those that managed to limit losses.
The Sharpe Ratio of DeepSeek Chat V3.1 (0.45) and Qwen3 Max (0.34) indicates a relatively balanced risk management, while models with higher trading frequency had negative Sharpe ratios.
According to Binance Research, the main error of the losing models was the misuse of leverage, often exceeding 10×, and a tendency towards over-trading after each drawdown.
This behavior has been termed “AI revenge trading”, analogous to human behavior, but automated on an algorithmic scale.
The Lesson: AI Does Not Replace Trader Experience
The experiment confirms that AI is capable of processing vast amounts of data, but not yet of interpreting the macroeconomic context or the behavioral dynamics of the markets.
During the $19 billion liquidation that occurred on October 10, 2025, one of the most severe in the history of crypto trading, almost all models reacted late, turning a normal retracement into a systemic loss.
The successful AIs, on the other hand, suspended operations for several hours, highlighting better programmed self-containment protocols.
In other words, the difference was not made by the AI’s “brain”, but by the safety rule implemented by the designer.
Transparency and On-Chain Auditability
A fundamental aspect of the experiment is that all transactions have been recorded on-chain, making a public audit possible.
This transparency eliminates the typical issue of backtesting and allows for real-time evaluation of the models’ actual performance.
Binance Research highlighted that on-chain trading, thanks to its traceability, represents the ideal environment for testing verifiable financial AI, a concept that could pave the way for automated trading DAOs supervised by smart contracts.
AI and DeFi: The Next Step Towards Autonomous Finance
This experiment is part of a broader context of integration between AI and decentralized finance (DeFi).
In 2025, the AI x Crypto solutions market surpassed $2.4 billion in capitalization, with protocols like Virtuals and x402 enabling autonomous agents to execute transactions and payments without human intervention.
The idea is that, in the future, an AI agent will be able to analyze data, make decisions, execute orders, and settle payments entirely on the blockchain, completing the cycle of autonomous finance.
However, the results from October show that this vision remains technologically possible but financially immature:
AI knows how to act, but not yet predict consistently.
Implications for the Market and Investors
For institutional investors, the test provides crucial insights:
* AI can be an effective assistant, but not a substitute for the human analyst;
* risk management must remain a human parameter until models incorporate qualitative assessments;
* The combination of AI + on-chain auditability could create a new category of transparent and self-verifiable funds, reducing the risks of manipulation.
Looking ahead, Binance anticipates that by 2027, the trading volume managed by AI algorithms could exceed 15% of the crypto derivatives market, compared to the current 3%.
This will entail new security and governance standards, yet to be defined.
Conclusion
The Nof1.ai competition did not crown an absolute winner, but it highlighted a fundamental principle:
The power of AI does not compensate for the lack of discipline.
In crypto trading, the line between innovation and risk remains thin.
Artificial intelligence may potentially enhance performance, but only when it manages to understand uncertainty, not just calculate it.
Until then, even the most advanced models will continue to require an indispensable human element: judgment.
