In the inaugural Alpha Arena Trading Competition, organised by Nof1, six advanced artificial-intelligence models were each allocated US $10,000 to trade cryptocurrencies autonomously between October 17 and November 3, 2025.
The models ranged from well-known names like ChatGPT, Gemini, and Claude Sonnet to the Chinese-built Qwen from Alibaba and DeepSeek.
Despite their AI-powered sophistication, most of the models experienced steep losses: ChatGPT ended with just ~$3,800 (-63%) and Gemini around ~$4,485 (-56%). Only Qwen and DeepSeek managed modest gains (Qwen ~+20%, DeepSeek ~+4%).
The competition highlighted that even top-tier AI models are not yet reliably predictive in the volatile crypto-trading environment, especially when using autonomous strategies without human oversight.
Why It Matters
- The event serves as a stress test for deploying AI trading models in one of the most unpredictable asset classes: crypto.
- It illustrates that large language models (LLMs) or AI systems trained on text/data may struggle when applied to high-speed financial trading, which demands adaptive numeric reasoning and real-time market reflexes.
- It underscores that hype around “AI trading” may need to be tempered with realistic expectations: just because a model can write text doesn’t mean it can trade crypto profitably.
- The contest may push more research into AI models specifically trained for financial markets rather than general-purpose LLMs.
Key Insights & Observations
- Model bias mattered: Some models showed systematic behaviour (e.g., mostly long positions, avoidance of shorts). In one case, Claude Sonnet was described as an “eternal optimist,” refusing to go short.
- Winning wasn’t about fancy tech: The two modest winners (Qwen, DeepSeek) employed simpler or specialised strategies rather than general “language model intelligence”.
- Market context played a role: The period covered included volatile crypto moves and the difficulty for a purely algorithmic model to adapt to changing conditions.
- This is an early stage: The organisers plan to expand future rounds (e.g., equities trading), indicating this was a prototype for much more advanced competitions.
Risks & Caveats
- The competition involved a small sample size (six models) in one segment of crypto trading; results may not generalise widely.
- Losses reflect short-term performance; a single time window in a volatile market is not definitive proof of model failure.
- Models were autonomous with no human intervention; real-world trading often includes human oversight, risk controls, and adaptive algorithms.
- The metric was profit/loss: other factors (risk-adjusted returns, drawdown, adaptability) are also important and not fully disclosed.
In the future:
- More contests with larger model pools, longer timeframes, and diversified asset classes.
- Research into specialised AI trading agents trained on numeric/market data rather than general language models.
- Increased scrutiny by traders, quant firms, and researchers around the suitability of AI in autonomous trading roles, particularly in crypto.
- Learning from failures: the market may shift from “AI will make us rich” to “AI needs domain-specific tailoring” when applied to trading.
FAQs
Q: What was the objective of the AI crypto-trading competition?
Six AI models were each given US$10,000 and tasked with autonomously trading several major cryptocurrencies over a fixed period (Oct 17 – Nov 3, 2025) to maximize returns.
Q: Which AI models participated?
Participants included ChatGPT, Gemini, Claude Sonnet, Grok, Qwen, and DeepSeek.
Q: How did they perform overall?
Most models lost significant portions of their capital (some over 50%). Only two models ended in profit: Qwen (~+20%) and DeepSeek (~+4%).
Q: What does this outcome tell us about AI in trading?
It suggests that current AI models, especially general-purpose ones, may struggle in highly volatile financial markets like crypto without specialist training or domain-specific design.
Q: Will there be more competitions like this?
Yes. The organisers have indicated that future rounds will include more models and potentially other asset classes such as equities.
