Benchmarks
348 sources · 1,258 models × 3,041 benchmarks · 158,957 scoresThe 206 GradeBench tasks where the student writes rather than draws, the complement of GradeBench-Diagram. Exact rubric match on those answers, from the same runs as the overall GradeBench score.
142nd strongest of 799 fitted benchmarks
Signal-to-noise +11.78 for the GradeBench family
Another benchmark represents this family in the fit.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Argon | medium | 57.7 | |
| 2 | Gemini 4 Carbon | medium | 56.4 | |
| 3 | Gemini 3.8 Flash | medium | 56.3 | |
| 4 | Claude Opus 5.5 | medium | 56.1 | |
| 5 | Claude Fable 5.1 | medium | 55.6 | |
| 6 | Claude Fable 5.2 | medium | 55.4 | |
| 7 | GPT-6.1 Sol | medium | 53.5 | |
| 8 | Gemini 4 Hydrogen | medium | 53.1 | |
| 9 | GPT-6 Astra | medium | 52.7 | |
| 10 | Claude Sonnet 5.5 | medium | 51.9 | |
| 11 | Muse Wonder | medium | 49.8 | |
| 12 | GLM-5.3 Flash | not reported | 47.6 | |
| 13 | Muse Spark 1.3 | medium | 44.0 | |
| 14 | Grok 4.7 | medium | 42.5 | |
| 15 | GPT-6 Luna | medium | 40.1 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).