Someone used Cloud Code to make a sailing simulation, then challenged other LLM holders to a race

Someone used Cloud Code to make a sailing simulation, then challenged other LLM holders to a race

summary

  • The Claude Code sim allows LLM certificate holders to manage virtual yachts using global status inputs every 5 seconds.

  • Gemini 4 Arjun won easily; GPT-6 Astra 2nd, Claude Opus stopped, and Quinn visited the beach.

  • The leading theory is that Gemini’s spatial awareness (relative to maps/earth) gave them an advantage over others.

If you gave a yacht to a Claude Opus 5.5, a GPT-6 Astra and a Gemini 4 Argon, and then asked them to sail across a body of water and back under the weather conditions, who do you think would win? While it would be a very expensive project (both in terms of getting the boats and feeding each LLM with credits), someone has done the next best thing. They used Claude Code to create a realistic regatta simulator, then tasked LLM frontier models with a virtual race.

Claude Code created a sailing field for four AI models to conquer

The results were widely varied

On the ClaudeAI subreddit, user abhi3188 showed what he made in Claude Code. It is a sailing simulator that allows an LLM student to take command of a virtual yacht and sail back and forth across a body of water. As for the sailors, abhi3188 chose a Claude Opus 5.5, a GPT-6 Astra, a Gemini 4 Argon, and a Qwen 3.5-9B, which they admitted they “threw in there for fun.”

Here’s how the regatta works:

Every 5 seconds, each model obtains the complete state of the world (23 prime numbers such as heading, speed, wind, location of the next mark, etc.) and determines the rudder and sails. There are no sailing tips. The simulation card pauses while thinking, so slow models are not penalized.

As for the race, Quinn came in last place because he refused to race and instead visited a nearby beach. Despite his contribution to programming the simulation card, Claude Opus 5.5 came in third place after stopping on his way home and running out of time. GPT-6 Astra took second place after completing a full cycle, and Gemini 4 Argon took an easy win, at a fifth of the token cost to Astra.

People have speculated about why Jiminy is such a superior sailor here, and the leading theory is that Google gave him superior spatial awareness because they also want to use him in Google Maps and Earth. Either way, I think I’ll enjoy more of these AI duels. Battlebots, anyone?

Leave a Reply

Your email address will not be published. Required fields are marked *