TangramPuzzle: leaderboard
TangramPuzzle is a multimodal benchmark for image-based compositional spatial reasoning using Tangram assemblies. The evaluated model performs outline prediction and end-to-end code generation from visual inputs. The pinned metric, Acc (%), measures accuracy against exact, machine-verifiable coordinate specifications. Source: BenchmarkList’s attributed result table.
Metric: Acc (%). Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 98.65 | #82 |
| 2 | GPT-5.2 | 77.54 | #115 |
| 3 | Claude Sonnet 4.5 | 51.2 | #148 |
| 4 | GLM-4.6V | 30.84 | #330 |
| 5 | InternVL3-78B | 29.19 | #361 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=tangrampuzzle · How It Works · Data refreshed daily, snapshot 2026-11-05.