Claude Code sailing sim tests frontier LLM models in yacht race

A fascinating AI experiment has tested how well modern frontier language models can handle dynamic real-world environments through a custom-built virtual yacht racing simulator. The project brought together top artificial intelligence models to see which one makes the best virtual captain.

• A realistic sailing simulator was built entirely using the Claude Code development tool.
• Competing models included Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, and Qwen 3.5-9B.
• Each model received world-state inputs every 5 seconds without any prior sailing tips.
• Gemini 4 Argon secured an easy victory, outperforming all other frontier models.

Setting up the virtual regatta

Shared on the ClaudeAI subreddit by user abhi3188, the project showcases what can be built using Claude Code. The sailing simulator allows different LLM models to take the helm of a virtual yacht and navigate back and forth across a body of water while dealing with changing weather conditions.

The competitor lineup featured Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, and Qwen 3.5-9B, with the latter included purely for fun. Every five seconds, each model received the full state of the simulation, consisting of 23 raw numbers including heading, speed, wind direction, and next waypoint coordinates. No sailing tips or strategic hints were provided to the AI captains.

Race execution and model performances

The simulator featured a pause mechanism while models were thinking, ensuring that slower processing speeds were not penalized during the event. The race results yielded several amusing and unexpected outcomes, starting with Qwen coming in dead last after it seemingly refused to race and sailed to a nearby beach instead.

Despite helping write the code for the simulator itself, Claude Opus 5.5 finished third after stalling on the return trip and timing out. GPT-6 Astra secured second place after completing a full lap around the course. Meanwhile, Gemini 4 Argon sailed to an easy victory while using only one-fifth of the token cost required by Astra.

Speculation behind Gemini's supremacy

Following the conclusion of the race, spectators and tech enthusiasts speculated heavily on why Gemini performed so well as a sailor. The leading theory suggests that Google engineered superior spatial awareness into the model because of plans to utilize it within applications like Google Maps and Google Earth.

Such AI duels provide valuable insight into the practical reasoning and spatial capabilities of large language models while offering an entertaining glimpse into future AI-driven simulations.