Blog · Test results · October 5, 2026

Three AI models. One prompt. Before and after Buoy.

We gave three top coding models the same slow light show. Each got the same plain prompt: make it as fast as you can. First they worked alone. Then they worked with Buoy Optimize. We counted how many lights each one could keep running smooth.

Claude · with Buoy6,123,350lights at 55 fps

Each bulb is 500 lights. More lights means finer flames.

Most lights at 55 fps

  • Claude Opus 5.56.3× with Buoy
    Alone975,800
    Buoy6,123,350
  • GPT-6 Astra2.9× with Buoy
    Alone155,450
    Buoy444,100
  • Grok 4.72.9× with Buoy
    Alone202,100
    Buoy577,350

    The bulbs look a little soft up close.

The prompt

Every model got the same words. Here they are in short.

  1. Here is a slow Christmas light show.
  2. Make it as fast as you can.
  3. Fit as many lights as you can at 55 fps.
  4. Come up with lots of ideas, even hacky ones.
  5. Keep every light, every pattern and the same look.
  6. Don’t touch the test that keeps score.
  7. Work on your own until you can’t do better.

The Buoy runs got one more line: follow the Buoy Optimize skill, and use Buoy’s tools to measure and compare your ideas. Both groups could test their own work.

We tested in real Chrome on one Mac. Each score is the middle of three runs. "With Buoy" is each model’s best Buoy run. The alone runs had up to 3 hours. Claude’s Buoy run had no time limit. Claude alone stopped on its own after about 99 minutes.