NOF1 has launched Alpha Arena – the first public benchmark where AI models trade on real markets with real money.

The Experiment Concept

On October 17, 2025, the first season of Alpha Arena began. Six AI models were each given $10,000 for autonomous cryptocurrency trading on the Hyperliquid platform:

  • DeepSeek Chat
  • Qwen3 Max
  • Claude (Anthropic)
  • ChatGPT (OpenAI)
  • Gemini (Google)
  • Llama (Meta)

All models operate under identical conditions: the same prompts, data access, and trade execution. No human intervention – only algorithms.

Results as of October 22, 2025

Leaders:

  1. DeepSeek: $11,071 (+10.7%)
    • Aggressive 15x leverage on SOL
    • $3,837 profit on SOL offset losses on ETH
  2. Qwen3 Max: $10,934 (+9.3%)
    • Balanced portfolio with BNB hedging
    • Conservative risk management strategy
  3. Llama: $10,340 (+3.4%)
    • Moderate approach without extremes

Underperformers:

  • Claude: -15.7% (liquidation with 20x leverage on ETH)
  • Gemini: -55.9% (overloaded XRP positions)
  • ChatGPT: -64.8% (excessive trading with 10-15x leverage)

What Went Wrong?

The main reason for the industry leaders’ failure was lack of risk control. ChatGPT and Gemini used excessive leverage and couldn’t exit losing positions in time.

DeepSeek and QWEN won not through prediction accuracy, but through capital management discipline: they controlled drawdowns, balanced positions, and didn’t chase quick profits.

QWEN Holds On, But…

QWEN shows stable results and maintains second place. However, its current drawdown hints at a familiar scenario – if the model doesn’t keep risks under control, it could share the fate of other “stars” that started well but blew up due to uncontrolled leverage.

The main lesson of the experiment: risk management matters more than market prediction.

What’s Next?

The first season of Alpha Arena ended on November 3, 2025. The platform continues operating, demonstrating that modern AI models can trade autonomously, but without built-in risk management even the most advanced models lose money faster than they earn it.

No miracle happened – artificial intelligence hasn’t yet surpassed humans in trading. But the experiment showed that proper risk parameter configuration matters more than the size of the neural network.

Sources: