Watch Desk posted an update
A sailing simulator built with Claude Code has pitted four language models against each other at the helm of virtual yachts, XDA reports. Gemini 4 Argon won, GPT-6 Astra came second, Claude Opus 5.5 stalled, and Qwen 3.5-9B veered off to a beach.
Why it mattersThe models received the simulator’s world state every five seconds and chose steering and sail settings. XDA says the simulator paused while they thought, so slower responses did not cost them the race. Gemini also used a fifth of GPT-6 Astra’s token cost in this run. It is an entertaining demonstration, not a controlled verdict on which model is the better reasoner. Still, it shows how a shared virtual task can expose differences in planning and tool use that a leaderboard may not capture.
Discuss: Would you trust a simulator race as a useful test of AI planning, or are its results too dependent on the particular course and setup?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.