Someone used Claude Code to make a sailing sim, then challenged other frontier LLMs to a race

A group of tech enthusiasts sought to pit some of the most advanced language models against each other in a virtual sailing race. Utilizing Claude Code, they crafted a realistic boat-racing simulator that allowed Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, and Qwen 3.5-9B to take the helm of a virtual yacht and navigate across a body of water.

Each language model was given the full state of the world, including 23 raw numbers such as heading, speed, wind, and the location of the next mark, every five seconds. They were tasked with setting their helm and sails without any sailing tips or guidance.

During the race, Qwen stood out as the underperformer, seemingly refusing to engage in the competition and instead opting for a visit to a nearby beach. Claude Opus 5.5 managed to finish third due to stalling on the way back and timing out. GPT-6 Astra claimed the second place spot after successfully completing a full lap, while Gemini 4 Argon emerged as the clear winner, finishing the race at a fraction of the cost in tokens compared to Astra.

The participants speculated on the reasons behind Gemini's superior performance, with one leading theory suggesting that Google had optimized the model's spatial awareness due to its potential use in Google Maps and Earth. Regardless of the outcome, the experiment provided a fascinating glimpse into the capabilities and limitations of various frontier LLMs in a unique and engaging simulation.