Worlds
Image-to-world builds. 3D worlds you can open in your browser.
An AI coding agent gets one AI-generated picture and builds it as a Three.js world, one run per build, with no human edits. In some tests a fresh session then improves a build.
These are not CmdBench benchmark runs: no checks, scores or verdicts.
-
Reference picture, AI-generated. Every run got this picture. GPT-6.2, same picture
OpenAI's GPT-6.2 Sol and GPT-6.2 Astra in Codex CLI each built the Lakeside build picture as a Three.js world on low, medium, high and xhigh effort, from the same prompt and the same copy of Three.js as the Claude runs. One run each, with no human edits.
-
GPT-6.2 Sol, low -
GPT-6.2 Sol, medium -
GPT-6.2 Sol, high -
GPT-6.2 Sol, xhigh -
GPT-6.2 Astra, low -
GPT-6.2 Astra, medium -
GPT-6.2 Astra, high -
GPT-6.2 Astra, xhigh
-
-
Reference picture, AI-generated. Every run got this picture. Gemini 4 Hydrogen, same picture
Google's Gemini 4 Hydrogen in Antigravity CLI built the Lakeside build picture as a Three.js world on high and xhigh effort, from the same prompt and the same copy of Three.js as the Claude, GPT and Muse Ember runs. One run each, with no human edits.
-
high -
xhigh
-
-
Reference picture, AI-generated. Every run got this picture. Claude Fable 5.2, same picture
Claude Fable 5.2 in Claude Code built the Lakeside build picture as a Three.js world on low, medium and high effort, from the same prompt and the same copy of Three.js as the Claude Opus 5.5 runs. One run each, with no human edits.
-
low -
medium -
high
-
-
Reference picture, AI-generated. Every run got this picture. Haiku build, Opus fix
Claude Haiku 5.5 in Claude Code built one AI-generated picture as a Three.js world on low and on medium effort. Then a fresh Claude Opus 5.5 session on xhigh improved a copy of each build. One run each, with no human edits.
-
low -
medium -
low + fix -
medium + fix
-
-
Reference picture, AI-generated. Every run got this picture. Japanese garden
Claude Opus 5.5 in Claude Code built a new AI-generated picture, a miniature Japanese garden, as a Three.js world on medium, high and xhigh effort. The prompt is Lakeside build's with the garden's scene nouns, the same sun and wind sliders and the same copy of Three.js. One run each, with no human edits.
-
medium -
high -
xhigh
-
-
Reference picture, AI-generated. The run got this picture. Muse Ember, same picture
Meta's Muse Ember in OpenCode built the Lakeside build picture as a Three.js world on high effort, from the same prompt and the same copy of Three.js as the Claude and GPT runs. One run, with no human edits.
-
high effort
-
-
Reference picture, AI-generated. Every run got this picture. Other models, same picture
GPT-6.1 Terra, GPT-6.1 Sol and GPT-6 Astra in Codex CLI each built the Lakeside build picture as a Three.js world on high effort, from the same prompt and the same copy of Three.js as the Claude runs. One run each, with no human edits.
-
GPT-6.1 Terra -
GPT-6.1 Sol -
GPT-6 Astra
-
-
Reference picture, AI-generated. Every run got this picture. Build on low, fix on xhigh
Claude Opus 5.5 in Claude Code built one AI-generated picture as a Three.js world on low effort. Then a fresh xhigh session improved a copy of that low build, and another one a copy of the medium build from Lakeside build. One run each, with no human edits.
-
low -
low + fix -
medium + fix
-
-
Reference picture, AI-generated. Every run got this picture. Sonnet build, Opus fix
Claude Sonnet 5.5 in Claude Code built one AI-generated picture as a Three.js world on low and on medium effort. Then a fresh Claude Opus 5.5 session on xhigh improved a copy of each build. One run each, with no human edits.
-
low -
medium -
low + fix -
medium + fix
-
-
Reference picture, AI-generated. Every run got this picture. Lakeside build
One picture, built as a 3D world by Claude Opus 5.5 in Claude Code at three effort levels: medium, high, xhigh. One run each.
-
medium -
high -
xhigh
-