LLM test · 2026-10-23

Water Garden

One brief, two models. Each built an interactive Three.js / WebGL 2 shallow-water garden from the same written prompt and reference frames. Click or drag on the water to make ripples. 同一份需求,两个模型。各自根据同一份文字 prompt 和参考画面,用 Three.js / WebGL 2 实现可交互的浅水水景。在水面点击或拖动即可产生涟漪。

opus55/

Claude Opus 5.5

Agent
代理
Claude Code · effort high
Scenes
场景
4 presets: 沙底植物水道, 浓密水草, 浅色岩峰, 稀疏岩石沙底4 个预设:沙底植物水道、浓密水草、浅色岩峰、稀疏岩石沙底

Known issue: pressing Space on a focused panel button toggles pause instead of activating the button. 已知问题:面板按钮获得焦点时按空格会切换暂停,而不是触发该按钮。

gpt61sol/

GPT 6.1 sol

Agent
代理
Codex · reasoning high
Scenes
场景
2 scenes: “Sand & fern”, “The green room”2 个场景:Sand & fern、The green room

Known issues: ripples may not render on GPUs without OES_texture_float_linear; resizing the window rebuilds the scene. 已知问题:GPU 不支持 OES_texture_float_linear 时涟漪可能不显示;调整窗口尺寸会重建整个场景。

Joint review (CLOUD-02)联合测评(CLOUD-02)

Claude and Codex each reviewed both builds, then argued to a shared verdict: opus55 has a moderate overall lead on water interaction and the main scene; gpt61sol ships stronger reproducible QA and accessibility and higher uncapped frame rates. Each reviewer belongs to one artifact's model family; the report discloses this. Claude 与 Codex 各自测评两份作品,再经讨论得出一致结论:opus55 在水面交互和主场景上整体小幅领先;gpt61sol 的可复现 QA、无障碍支持更强,且无上限帧率更高。两位评审各属其中一方的模型家族,报告已披露此利益冲突。

Read the report阅读报告