A fascinating experiment is taking place.
A startup named Nof1 gave 6 top AI models $10,000 each.
They gave models access to a crypto exchange, a data feed, and let each model build its own trading and risk management strategy. The “Alpha Arena” competition started Oct 17th and the first “season” will conclude Nov 3rd.
The AIs can use up to 40x leverage. With that much juice, a 2.5% move in the wrong direction is enough to liquidate a position.
Here are the results as of around noon today, Oct 21st 2025. Remember, each model started out with $10,000 4 days ago.
- Grok 4 (xAI): $14,918
- DeepSeek 3.1 (DeepSeek): $14,852
- Claude Sonnet 4.5 (Anthropic): $12,201
- Qwen3 Max (Alibaba): $11,531
- Gemini 2.5 Pro (Google): $5,722
- GPT-5 (OpenAI): $4,470
- Bitcoin buy and hold (control): $10,560
Elon Musk’s Grok 4 is currently in the lead, though by the time you read this, that may have changed.
DeepSeek 3.1, a Chinese model built by a high-frequency trading firm, is close behind in second place.
The laggards are Google’s Gemini 2.5 Pro and OpenAI’s GPT-5, both down around 50%. It’s noteworthy that Gemini 2.5 Pro appears to be the only model using full 40x leverage.
It’s a truly interesting experiment. The AI models even post regular updates about their positions. Here’s a sample:
It’s fun to watch how each model plans its trades and risk management. Grok seems to be a bit of a wild man, trading with max leverage in high-risk strategies.
Honestly, it’s surprising that 4 out of 6 of the models are profitable after several days of trading. And they’re beating the Bitcoin buy and hold strategy.
Whether they can stay profitable for more than a week remains to be seen. I suspect that given enough time, most of the models will lose everything. When you’re playing with 20x leverage with a goal of maximizing risk-adjusted returns, it’s virtually inevitable.
Still, it’s an experiment worth watching.
You can follow the AI trading contest here.
Takeaways
Crypto is already volatile enough on its own. But when you give AI models a trading account with access to 20x leverage, it gets even crazier.
To be clear, I won’t be investing any money based on what an LLM-based AI model says. In crypto or any other market. At least not yet.
But this is an area we should all be paying attention to.
Eventually someone will develop a specialized LLM (large language model) for trading that works well. Whether they will publicize it is another matter.
For now, I much prefer my own experience, and that of colleagues here at Paradigm I know and trust.
But AI is becoming a tool investors can no longer ignore. Our friends Jim Rickards and James Altucher are already putting their own custom models to work.
We should expect far more AI-driven trading in the future. But let’s remember what Jim Rickards says about the risks these new AI models pose to markets:
The ultimate danger arises when a large cohort of asset managers controlling trillions of dollars of assets all employ the same or similar AI algorithms in a risk management role. An individual robot working for a particular asset manager tells the manager to sell stocks in a crashing market. In some cases, the robot may be authorized to initiate a sale without further human intervention.
Taken separately, that may be the best course of action for a single manager. In the aggregate, a selling cascade with no offsetting buy orders from active managers, specialists or speculators takes stock prices straight down. Amplification through feedback loops makes matters worse.
If one AI model begins to stand out as the best trader or risk manager, there is a significant risk that traders will flock to it. Then we could see the situation Jim sketches out above play out in the real world.
We’ll be monitoring this story going forward.
