GPT-6.2, same picture

One picture, eight 3D worlds. GPT-6.2 Sol and GPT-6.2 Astra in Codex CLI at four effort levels.

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.

This is not a CmdBench benchmark run: no checks, scores or verdicts.

AI-generated picture of a miniature island: a wooden cabin with a stone chimney, a jetty and rowboat, a yellow crane over a half-built timber frame, lumber stacks, rocks and pine trees, in warm late-afternoon light.
Reference picture, AI-generated. Every run got this picture and the same prompt.

The eight worlds

One run per model and effort level. Each world is served exactly as the model wrote it, with no fixes, and opens full screen in this tab.

One difference from the Claude runs: Codex CLI can't be limited to file tools, so these runs could also run commands in a sandbox (their own folder, no network, no browser). The logs show none did; each only wrote files.

Costs are computed estimates at API list prices: Codex CLI reports no cost, so each is the run's own token counts at OpenAI's list prices of 17.11.2026.

How it was made

Picture
AI-generated. The exact PNG every run got
Prompt
The same for every run: read the prompt (TASK.md)
Models
GPT-6.2 Sol (gpt-6.2-sol), GPT-6.2 Astra (gpt-6.2-astra)
Agent
Codex CLI 0.169.0
Effort levels
low, medium, high and xhigh for both models; one run each
Tools
Codex CLI can't be limited to file tools, so these runs could also run commands in a sandbox: their own folder only, no network, no browser. The logs show none did; each only wrote files. No run saw its world running.
Time limit
60 minutes per run
Three.js
0.187.0
Human edits
None. Each world is served exactly as the agent wrote it, with no fixes.
Run dates
17.11.2026 and 18.11.2026