Four AI models race sailboats in a Claude Code-built sim
Reddit user abhi3188 built a sailing simulator with Claude Code and raced four language models in it. In XDA's account, Gemini 4 Argon won at a fifth of GPT-6 Astra's token cost and Claude Opus 5.5 stalled. How the test worked, what it can't tell you, and a prompt to build your own.
Updated November 20, 2026: Expanded with how the race was set up, what such a test measures, its limits, and a prompt for building a similar sim.
A Reddit user who posts as abhi3188 built a sailing simulator with Claude Code. In it, a language model takes the helm of a virtual yacht and has to sail across a stretch of water and back. They then raced four models against each other. Simon Batt of XDA Developers wrote up the result on November 7, 2026, quoting the original post on the ClaudeAI subreddit. We could not open the Reddit thread itself when updating this article, so everything below comes from XDA's report and its quotes from the builder, not from the post, its comments or any logs.
How the race worked
According to the post XDA quotes, every five seconds each model received the full state of its boat and the course as 23 raw numbers. These included heading, speed, wind and the position of the next mark. The model then set its helm and sails. The builder says the models got no sailing tips. The simulation paused while a model was thinking, so slower models were not penalised for taking longer to answer.
The entrants were Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and Qwen 3.5-9B. The builder said they threw the Qwen model in for fun. The model names here are as XDA reports them.
The result
- Gemini 4 Argon won easily, at about a fifth of GPT-6 Astra's token cost. XDA gives the ratio only, not token counts or dollar figures.
- GPT-6 Astra came second after completing a full lap.
- Claude Opus 5.5, which had helped write the simulator, stalled on the way home and timed out, finishing third.
- Qwen 3.5-9B came last, in XDA's words seemingly refusing to race and visiting a nearby beach instead.
XDA notes that commenters speculated Gemini's edge came from spatial awareness Google may have built for products like Maps and Earth. That is a theory from the thread, not something the test shows.
What a test like this can and can't tell you
This is one race on one simulator, written up from a social media post, so in our judgement it is closer to an anecdote than a benchmark. XDA's report doesn't say how many runs were made, how long the race took, how the course or wind was set, how each model was prompted, or how long the timeout was. It also doesn't explain why Opus 5.5 stalled or why Qwen headed for the beach. With a single run, one stall is enough to change the finishing order. Still, the format is easy to read, in our view: the input is a fixed set of numbers, the output is helm and sails, and a failure is a boat you can watch going the wrong way.
Build your own
XDA's report does not link to the simulator's code, so a prompt is the closest you can get. The prompt below keeps the rules the builder described: raw numeric state every five seconds, no sailing tips and a paused clock while each model thinks. Everything else in it, such as the course, reply format and logging, is our own design choice, so a sim built from it is a different test, not a replication. Calling several commercial models will need API keys and will cost money.
Try it with your AI. Paste this into Claude Code in an empty folder to scaffold a small control benchmark:
Build a sailing simulator with a simple visual view. Model a single yacht with heading,
speed, rudder angle and sail trim, and wind. Add a course with marks to round. Every 5
simulated seconds, pause the simulation and send the full world state to a model as raw
numbers only, with no sailing advice. The model replies with helm and sail settings
within fixed ranges; reject and log anything else. Keep the clock paused while the model
is thinking. Support several model backends behind one interface, read API keys from
environment variables, and log every request, response, token count and boat position so
a race can be replayed. Add a timeout for a model that stops making progress, and record
where it was when it timed out.Our take
In our view, small physical sims like this are some of the most readable model tests around: you don't need to trust a leaderboard when you can watch a boat drift onto a beach. The obvious next step for anyone repeating it is many races with varied wind and published logs. That would show whether Gemini's win and Opus 5.5's stall are patterns or one-off results. It is also a reminder that the model that writes the simulator is not necessarily the one that sails it best.
Built by Reddit user abhi3188
This story first appeared in the claudefox daily briefing for November 8, 2026.