Same jelly-card brief: Opus 5.5 builds real soft-body physics, GPT-6 fakes it with curves
In a comparison video, Opus 5.5 uses 81 points on springs stepped 240 times a second to make volume-preserving jelly, while GPT-6 Astra fakes the deformation with animation curves, taking 1h40m versus 43 minutes.
I gave Claude Opus 5.5 and ChatGPT-6 Astra the same brief, word for word: make these portfolio cards jelly. Astra shipped in 43 minutes. Claude took 1h 40m. Claude built a real soft body: 81 points on springs, stepped 240 times a second. The jelly keeps its volume, so a squeezed card gets thicker. Astra faked it with animation curves, and I love that its own README says so. Both get the same 14 moves in this clip. You can fake a bend. You can't fake the splat.