PAGE Bench - Final Result Score: leaderboard
Final-result score of PAGE Bench (Wei et al., 2026): the figure an agent constructs in GeoGebra for each of 463 geometry problems is compared with the reference construction. A rule-based point and command match counts 0.3, and a Gemini-2.5-Pro judge's task completion, visual similarity and geometric logic count 0.3, 0.2 and 0.2, on a 0-100 scale; higher is better.
Metric: Final Result Score (0-100; 0.3 point and command match + 0.3 task completion + 0.2 visual similarity + 0.2 geometric logic, the last three from a Gemini-2.5-Pro judge). Source: arxiv.org. Saturation forecast: Around 2028. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 16.31 | #91 |
| 2 | Qwen 3.6 Plus | 11.72 | #217 |
| 3 | GPT-5.4 | 9.96 | #121 |
| 4 | Claude Sonnet 4.6 | 8.65 | #158 |
| 5 | GLM-4.5V | 7.27 | #739 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=page-bench-final-result-score · How It Works · Data refreshed daily, snapshot 2026-11-12.