Atomic Chat compared four frontier models on a physics demo task, measuring token cost and output quality.

• Fable 5 produced the highest quality physics demos but cost 6x more than Opus 4.8.
• GPT 5.5 was closest to Fable 5 in quality and cheaper, while GLM 5.2 was cheapest but lower quality.
• The benchmark used three specific HTML5 canvas scenes: train derailment, car collision, and monster truck crushing cars.

how to reproduce · Give each model the same prompt: 'Build three self-contained HTML5 canvas scenes with real physics demos: (1) A train derailing off a broken bridge into the water, (2) Two cars jumping off ramps and colliding mid-air over a canyon, (3) A monster truck crushing a row of parked cars.' Record token usage and cost for each model, then evaluate output quality.