Gemini 4 Argon wins virtual yacht race against other frontier models

The virtual yacht race

A user on the ClaudeAI subreddit recently conducted a unique experiment to see how frontier AI models handle navigation. They used Claude Code to build a realistic boat-racing simulator where different models could take the helm of virtual yachts. The goal was to see which AI could successfully sail across a body of water and back while accounting for specific weather conditions.

The competition featured four distinct models: Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5, and Qwen 3.5-9B. While the user admitted that Qwen was included mostly for fun, the other three represented the current state of high-end artificial intelligence. The experiment provided a way to test spatial reasoning without the massive expense of physical boats or real-world logistics.

As reported by XDA, the results were notably varied among the participants. While some models managed to complete the course, others struggled with the basic physics or logic required to stay on track. The test served as a practical benchmark for how these systems process continuous inputs to make sequential decisions in a simulated environment.

How the simulator works

The simulator was designed to provide the AI models with a constant stream of data to inform their movements. Every 5 seconds, each model received the full state of the world, which consisted of 23 raw numbers. These data points included critical information such as the boat's current heading, its speed, the wind direction, and the location of the next mark.

To ensure the models had enough time to process this information, the simulator was programmed to pause while each AI was thinking. This design choice meant that slower models were not penalized for taking longer to calculate their next move. Each model was tasked with setting its own helm and sail positions based solely on the numbers provided, without receiving any specific sailing tips or guidance.

By stripping away conversational help and providing only raw data, the experiment forced the models to interpret the environment independently. The AI had to translate numbers into a coherent understanding of motion and wind resistance. This setup effectively turned the models into autonomous pilots navigating a digital sea.

Model performance results

Gemini 4 Argon emerged as the clear winner of the virtual race. It navigated the course successfully and did so with significant efficiency, using only 1/5th of the tokens compared to GPT-6 Astra. This suggests that Gemini was able to solve the navigation problem with much higher precision and less computational overhead than its competitor.

GPT-6 Astra took second place, managing to complete a full lap of the water. However, it was less efficient than Gemini in terms of the resources it consumed to reach the finish line. The performance gap between these two leading models highlighted differences in how they prioritize and process spatial information during complex tasks.

The other participants struggled significantly with the simulation's requirements. Claude Opus 5.5, which helped code the simulator itself, finished in third place after it stalled on the way home and timed out. Meanwhile, Qwen 3.5-9B came in last, as it seemingly refused to participate in the race and instead navigated toward a nearby beach.

The spatial awareness theory

The disparity in performance has led to interesting theories regarding why Gemini 4 Argon excelled so well. One leading theory suggests that Google has provided the model with superior spatial awareness. This capability is likely a byproduct of the work done to integrate AI into products like Google Maps and Google Earth.

Because these applications require a deep understanding of 3D space and geography, the underlying model may be better at