四个模型同题做骑车 three.js 动画对比
同一句提示词让 GPT-6.1 Sol、Opus 5.5、Sonnet 5.5、GPT-6 Astra 各做一段骑车循环动画，同屏对比运镜与光照。
Have they buffed GPT-6.1 Sol? I gave it one line: "Create a man driving a bycicle three.js detailed loop in single html file." Typo included. It came back with a full scene. Articulated legs, pedals and chain that actually move, a 16-second seamless loop, morning and dusk lighting, orbit and zoom, pause and speed controls. One offline HTML file. So I ran the same idea through three more, this time with a full written brief: Claude Opus 5.5 (xhigh, Claude Code): a cobalt-jersey rider on a burgundy steel bike through the Cape Winelands, vineyards and Cape Dutch gables. 30-second loop. Claude Sonnet 5.5 (xhigh, Claude Code): a bearded rider in rust on a teal frame, golden hour on a Mediterranean back road. 20-second loop. GPT-6 Astra (xhigh, Codex): a coastal ride with a sailboat on the water. 12-second loop. The video runs all four on one clock with the same camera sway, so you're comparing like with like. What I noticed: Sol got the least instruction and still delivered a complete, working scene. Opus has the richest setting. Sonnet has the strongest lighting. Astra has the cleanest composition. On "buffed": OpenAI did make Astra and 6.1 Sol about 50% faster on Monday, Day 1 of Tibo's 28 days. That's speed, not smarts. Whether the output got better is your call, so open them and spin them yourself: GPT-6.1 Sol: http://dewaldt.link/bicycle-gpt61-sol Claude Opus 5.5: http://dewaldt.link/bicycle-opus55-xhigh Claude Sonnet 5.5: http://dewaldt.link/bicycle-sonnet55-xhigh GPT-6 Astra: http://dewaldt.link/bicycle-astra-xhigh Not a lab benchmark. Sol ran in a Hermes chat from a one-liner; the other three got the same detailed brief. Rendered headless without a GPU, so phones and real GPUs will run smoother. Which one would you ride?