Atlas

Benchmarks

GradeBench

GradeBench (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.

Models
15
Top score
57.3
Leader
Claude Fable 5.2
Updated
Nov 16, 2026
Resolution
~2 mo
Sharp
Saturation
~Jul 2028
Projected
Signal
82ndpercentile
197 benchmarks from 0.8 to 3.5227 benchmarks from 3.5 to 6.2132 benchmarks from 6.2 to 8.996 benchmarks from 8.9 to 11.668 benchmarks from 11.6 to 14.326 benchmarks from 14.3 to 17.013 benchmarks from 17.0 to 19.78 benchmarks from 19.7 to 22.39 benchmarks from 22.3 to 25.03 benchmarks from 25.0 to 27.76 benchmarks from 27.7 to 30.41 benchmark from 30.4 to 33.14 benchmarks from 33.1 to 35.85 benchmarks from 35.8 to 38.51 benchmark from 38.5 to 41.22 benchmarks from 43.9 to 46.61 benchmark from 62.7 to 65.4

142nd strongest of 799

Signal-to-noise +11.78

Difficulty
74thpercentile
1 benchmark from 5.6 to 14.51 benchmark from 23.3 to 32.22 benchmarks from 32.2 to 41.11 benchmark from 41.1 to 49.92 benchmarks from 49.9 to 58.89 benchmarks from 58.8 to 67.67 benchmarks from 67.6 to 76.518 benchmarks from 76.5 to 85.424 benchmarks from 85.4 to 94.241 benchmarks from 94.2 to 103.131 benchmarks from 103.1 to 112.040 benchmarks from 112.0 to 120.849 benchmarks from 120.8 to 129.776 benchmarks from 129.7 to 138.560 benchmarks from 138.5 to 147.4117 benchmarks from 147.4 to 156.3126 benchmarks from 156.3 to 165.1121 benchmarks from 165.1 to 174.038 benchmarks from 174.0 to 182.820 benchmarks from 182.8 to 191.710 benchmarks from 191.7 to 200.61 benchmark from 200.6 to 209.44 benchmarks from 209.4 to 218.3

209th hardest of 799

Capability needed 164.0

Organizations
15 dated models · 6 frontier modelsClosed weightsOpen weights
Data: Pensiveatlas.village.dev

GradeBench leaderboard

Related benchmarks

Loading Atlas data…