Can AI models trade better with just charts or full market data?

We ran a practical test on this question. Four leading frontier AI models—along with their vision-only variants—were given real capital to execute trades on ETH on Aerodrome, a popular DEX.

The results told an interesting story:

Sonnet 4.5 with full data came out on top, delivering +0.06% returns. Not massive gains, but it outperformed every competitor.

Gemini 3 Pro Vision followed close behind, managing -0.20%—impressive for a vision-only model working with just chart data.

Grok-4 with data access fell further behind at -0.99%.

Where GPT-5.2 landed? It underperformed significantly.

The takeaway? It's not straightforward. While data-rich models theoretically have an edge, vision-only approaches sometimes held their own. The gap between models was surprisingly narrow in some cases—suggesting that trading success depends on more than just information access. Execution logic, risk management, and model architecture clearly matter.