SAGE: leaderboard
Grading of handwritten maths coursework (automata, linear algebra, PDEs, AP calculus) against instructor rubrics, where the model picks which point deductions apply; Vals AI; top model 56.1%.
Metric: Accuracy (%). Source: www.vals.ai. Status: saturation imminent. 80 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.7 (Max) | 56.1 |
| 2 | Claude Opus 4.8 (Max) | 54.79 |
| 3 | Kimi K3 | 54.26 |
| 4 | GPT-5.6 Sol (Max) | 52.56 |
| 5 | Qwen 3.8 27B (xHigh) | 52.4 |
| 6 | Claude Fable 5 (Max) | 51.89 |
| 7 | Gemini 3 Flash (Preview) (High) | 51.85 |
| 8 | Claude Opus 4.6 (Adaptive Reasoning, Max Effort) | 51.58 |
| 9 | GPT-5.5 (xHigh) | 51.53 |
| 10 | Qwen 3.8 Max | 51.25 |
| 11 | GPT-5.4 Mini (xHigh) | 50.81 |
| 12 | MiniMax-M3 | 50.57 |
| 13 | Kimi K2.6 | 50.22 |
| 14 | Gemini 3.5 Flash (High) | 49.88 |
| 15 | Kimi K2.5 (Thinking) | 49.87 |
Interactive version: theaggregate.ai/benchmark?slug=sage · How It Works · Data refreshed daily, snapshot 2026-09-05.