MBABench 2.0 — Numerical Accuracy (Medium or Harder)
Professional Work · 2026-10-10
Share of the ninety Medium, Medium-Hard and Hard tasks for which every numerical answer matches the expert key. The published Overall accuracy excludes the eleven Easy tasks and does not apply the workbook-quality gate.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 58.9 |
| Claude Opus 5 | 54.4 |
| GPT-6 Astra | 54.4 |
| Grok 4.6 | 45.6 |
| Kimi K3 | 44.4 |