ShapeCodeBench: leaderboard

ShapeCodeBench (2026) tests perception-to-program reconstruction: from an image of filled and outlined circles and squares, a model writes a program in a small drawing language that is re-rendered and compared with the target. The score is mean foreground IoU in percent on the frozen 150-image eval_v1 split, three difficulty tiers; higher is better.

Metric: Mean foreground IoU (%) between the re-rendered and the target shape pixels on the 150-image eval_v1 split (50 per difficulty tier); the model writes a program in a four-primitive drawing DSL that is re-rendered and compared with the target raster, zero-shot, GPT-5.5 through the Codex CLI and Claude Opus 4.7 through Claude Code at the stated effort; a parse failure scores 0; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.5xHigh86.5#54
2Claude Opus 4.7Max46.1#85

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=shapecodebench · How It Works · Data refreshed daily, snapshot 2026-11-21.