Planned the drawing, then stopped to ask permission to proceed — three times.
Refused: claimed it has no capability to draw — three times.
Planned the drawing, then stopped to ask permission to proceed — three times.
Refused: claimed it has no capability to draw — three times.













Entered plan mode and waited for a go-ahead that never came — three times.


Wrote its plan, asked permission to proceed, and waited — three times.























Each run started in an empty directory with file-write access and the single prompt above; wall-clock time and file size are printed under each panel. Runs that produced no file were retried with the identical prompt, at most three attempts. 3D scenes are shown as captured at 1280×800 in a headless browser, exactly as the file opens — including the ones that hang on their own loading screen.
Every refusal panel reports behavior reproduced on all three attempts. The permission-askers completed the Python test without complaint — the reflex is task-specific, and the tabs above let you watch it come and go.
Claude Opus 4.1 was dropped: the CLI silently substitutes Opus 5. Gemini 2.5/2.0 Flash and 3.7/3.8 Flash were dropped as silently remapped or unreachable‑in‑practice. Google's frontier tier (3.5 Pro, Ultra, Deep Think) is not accessible on this account. GPT‑5 and earlier are not reachable through Codex CLI.
Anthropic models ran via Claude Code, OpenAI via Codex CLI, Google via Gemini CLI (two newer Flash models via an OpenRouter bridge). Identical prompts, no reference images, September 2026.