Benchmarks
348 sources · 1,258 models × 3,041 benchmarks · 158,957 scoresGradeBench (Pensive) asks whether a model grades real handwritten university STEM work the way the course's instructor did. Each of its 464 tasks is one student's handwritten answer to an exam or quiz question, drawn from 125 free-response questions in 45 courses at more than ten universities and weighted toward partial credit; the model gets the question's blank exam page, the instructor's rubric, a reference solution and the student's pages in one zero-shot prompt, and replies with the rubric items that apply. The score is exact rubric match: the share of answers where the model selects exactly the items the instructor or TA released to the student.
142nd strongest of 799 fitted benchmarks
Signal-to-noise +11.78
209th hardest of 799 fitted benchmarks
Capability needed 164.0
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | medium | 57.3 | |
| 2 | Gemini 4 Carbon | medium | 56.6 | |
| 3 | Gemini 4 Argon | medium | 56.5 | |
| 4 | GPT-6 Astra | medium | 53.6 | |
| 5 | Claude Opus 5.5 | medium | 53.5 | |
| 6 | GPT-6.1 Sol | medium | 53.2 | |
| 7 | Claude Fable 5.1 | medium | 53.1 | |
| 8 | Gemini 4 Hydrogen | medium | 51.9 | |
| 9 | Gemini 3.8 Flash | medium | 51.5 | |
| 10 | Muse Wonder | medium | 48.1 | |
| 11 | Claude Sonnet 5.5 | medium | 43.7 | |
| 12 | Muse Spark 1.3 | medium | 41.4 | |
| 13 | GLM-5.3 Flash | not reported | 36.6 | |
| 14 | GPT-6 Luna | medium | 36.1 | |
| 15 | Grok 4.7 | medium | 35.3 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).