Breaking

Developer Builds Sailing Simulator With Claude Code, Then Sets Four AI Models Loose at the Helm

November 6, 2026 at 19:46 EST

AI Agents
Software Dev & Coding

A developer says they used Claude Code to build a sailing simulator, then tested how Anthropic's Opus, OpenAI's GPT-6 Astra, Google's Gemini 4 Argon, and Alibaba's Qwen steer a boat using only raw world-state data refreshed every five seconds.

Closed-loop AI control · Sailing simulator experiment

Four AI models steer a simulated sailboat from raw numbers alone

A developer built the simulator with Claude Code and tested how Opus, GPT-6 Astra, Gemini 4 Argon, and Qwen handle the helm. Each model read only numerical state data refreshed every five seconds. No race results have been published yet.

AI models tested on the same helm task

numerical values sent per update

5 s

refresh interval for each decision

camera images or pre-written nav plans given

Scale of input per update

Each block = 1 numerical value

23 values

Heading, speed, wind conditions, and the position of the next mark are among the inputs described. The models had to interpret these numbers and choose steering inputs in real time, with no visual feed.

The control loop, repeated every update

1. Read 23 values

→

2. Interpret state

→

3. Issue helm input

→

4. Boat state changes

The new state feeds the next update, so each steering choice shapes the data the model sees next.

What the developer built

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.