User abhi3188 on the ClaudeAI subreddit built a custom boat-racing simulation using Claude Code and tasked four frontier large language models with piloting virtual yachts across a body of water. Gemini 4 Argon secured an easy win at a fraction of the token cost, while GPT-6 Astra finished second, Claude Opus 5.5 stalled out in third place, and Qwen 3.5-9B abandoned the race entirely to visit a nearby beach.
How the Virtual Yacht Race Functioned
The simulation operates by feeding each model 23 raw numbers every five seconds, detailing the exact state of the world including heading, speed, wind conditions, and the location of the next mark. The LLMs adjust their helm and sails based on these numerical inputs without any sailing tips or external assistance. The simulation pauses while each model computes its next move, ensuring that slower inference speeds do not penalize any specific competitor.
The lineup featured Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, and Qwen 3.5-9B. According to Xiang, Qwen was included purely for fun and ultimately refused to participate in the race parameters, steering toward a beach instead. Despite its foundational involvement in writing the underlying code via Claude Code, Claude Opus 5.5 stalled during the return leg and timed out, leaving it in third place. GPT-6 Astra successfully completed a full lap to take second place.
Gemini 4 Argon Claims Victory at Lower Token Cost
Gemini 4 Argon won the competition comfortably, finishing ahead of its rivals while consuming just one-fifth of the token cost required by GPT-6 Astra. Observers on social media speculated extensively about the performance gap. The leading theory suggests that Google optimized Gemini for superior spatial awareness because the architecture is designed to power Google Maps and Google Earth applications.
As xda-developers.com reported, the project highlights the emerging utility of AI coding agents like Claude Code in rapidly prototyping complex physics environments and custom test beds. The experiment also demonstrates distinct operational differences among frontier models when handling real-time spatial calculations and continuous state inputs.
Future AI Competitions Await
No official timeline or announcements have been released regarding follow-up events.
