The Great Crypto Challenge: GPT-5 vs. AI Models
What would happen if you entrusted $10,000 to GPT-5 for investment in cryptocurrencies ? What if you put the same amount in the hands of other advanced models and observed how they compete against one another? This intriguing concept was brought to life by Nof1. The results have proven to be both surprising and enlightening.
Six Rivals in the Crypto Arena
The masterminds behind Nof1 have launched Alpha Arena , an innovative benchmark that adapts its difficulty based on the intelligence of the AI involved. The fundamental aim is to measure the performance of six top-tier models in the realm of cryptocurrency investment using a budget of $10,000. The contenders include:
- GPT-5
- Gemini 2.5 Pro
- Claude Sonnet 4.5
- Grok 4
- DeepSeek Chat v3.1
- Qwen 3 Max
DeepSeek has turned its $10,000 into almost $20,000, while Qwen has managed $15,000. In contrast, both GPT-5 and Gemini 2.5 Pro have lost about 65% of their initial investments.
Triumph of DeepSeek and Qwen; GPT-5 and Gemini Struggle. After 11 days of competition, the outcomes are astonishing. The two robust models, DeepSeek and Qwen, have achieved significant returns: DeepSeek has seen a return of 97% (with peaks hitting 123%), while Qwen isn’t far behind at 53%. On the other hand, GPT-5 (-65.7%) and Gemini 2.5 Pro (-66%) are facing substantial declines, having lost two-thirds of their investments.
The performance indicators not only reflect gains and losses but also reveal the number of trades executed. GPT-5 engaged in 75 trades, while Gemini 2.5 Pro was even more active with 193 trades. Higher trading volumes, however, did not correlate with successful outcomes.
Evolving Crypto Fortunes
The models have the flexibility to invest in six major cryptocurrencies: Bitcoin , Ethereum , Dogecoin , Ripple , Solana , and BNB . They decide on their investment strategies, including the amount to invest, levels of leverage, and positions held. While many positions are typically maintained for a few hours, some last days.
Learning Through Competition
All models have been in competition since October 18 in what is termed the “first season” of this experiment, slated to conclude on November 3. As its creators explain, the findings from this initial round will provide foundational insights into how these models operate in financial markets.
Investing in cryptocurrency requires a keen understanding of market trends and data analytics.
Enhancing Investment Strategies
The primary goal of this competition is straightforward: maximize returns while minimizing losses (PnL). The objective of the first season is to refine the models based on the lessons learned through this exercise, paving the way for enhanced features and capabilities in future investments.
Innovative Algorithmic Trading
The strategies employed by these models might seem reckless for human investors due to their aggressive use of leverage , allowing for rapid profit maximization but also introducing extreme risks. These models adopt extraordinary leverage ratios, often reaching up to 20x or 25x, and can take short (bet against asset value) or long (bet for asset value) positions.
The structural mechanics of the benchmark experiment will evolve with ongoing seasons to yield richer results.
Rational Decision-Making in AI
To manage risks effectively, each model adheres to clear guidelines regarding risk limits, employing strategies like stop-loss orders. This disciplined approach allows them to maintain or adjust positions as necessary. Notably, these models engage in medium to low-frequency trading, making decisions over minutes or hours rather than in milliseconds, giving them sufficient time and information to act wisely.
A Cautionary Note for Investors
It’s crucial to recognize that this experiment is purely exploratory; it is beset with numerous financial caveats. The brief duration of the first season hinders comprehensive long-term evaluations. Moreover, the information available to the models is notably restricted, relying solely on numerical data such as average prices and trading volumes without incorporating broader market news.
DeepSeek demonstrates its reasoning for maintaining positions based on its analysis of current data.
Transparency in Decision-Making
The interface of the trading models allows users to examine how each AI rationalizes its trading decisions. By accessing the “Model Chat” section, users can explore both current and historical data considered by each model, gaining insights into their trading rationale.
The Road Ahead: Continual Improvement
While certain models may currently outperform others, Nof1’s creators emphasize that this experiment is not merely about identifying the “best” model. They are acutely aware of the inherent limitations of this inaugural season, including response bias and the brevity of the experiment duration. They plan to iterate through multiple seasons, enhancing decision mechanisms and data availability to measure performance consistency better. This study of AI’s efficacy in financial markets is genuinely fascinating.
Image | Aedrian Salazar
In Xataka | A country has undertaken the largest cryptocurrency experiment in the world: Bhutan and the 800,000 ID cards with Ethereum.