Gemini 4 Argon Wins LLM Sailing Race Built With Claude Code
Four models steered virtual yachts in a Reddit user's simulator, receiving 23 raw world-state numbers every 5 seconds with no sailing tips.
A Reddit user going by abhi3188 built a sailing simulator with Claude Code in which large language models helm virtual yachts across a body of water and back, and raced four models in it, according to xda-developers.com. Gemini 4 Argon won the race.
The four competitors were Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and Qwen 3.5-9B. Every 5 seconds, each model received 23 raw world-state numbers, including heading, speed, wind and the location of the next mark, and then set its helm and sails without any sailing tips.
The simulation pauses while a model thinks, so slower models are not penalised. Gemini 4 Argon finished first using about one-fifth of the tokens that GPT-6 Astra spent. GPT-6 Astra came second after completing a full lap.
Claude Opus 5.5 took third place after stalling on the return leg and timing out. Qwen 3.5-9B finished last, apparently refusing to race and visiting a nearby beach instead.
Why Gemini may have won
The leading theory offered in the article is that Google gave Gemini superior spatial awareness because it also wants to use the model in Google Maps and Earth.
Quick answers
Which AI model won the LLM sailing race?
Gemini 4 Argon won, using about one-fifth of the tokens GPT-6 Astra spent.
How did the models control the boats?
Every 5 seconds each model received 23 raw world-state numbers such as heading, speed and wind, and set its helm and sails without sailing tips.
Who built the sailing simulator?
A Reddit user named abhi3188 built it with Claude Code.