MBABench 2.0 — Hard Numerical Accuracy
Professional Work · 2026-10-10
All-answers-correct rate on the twenty Hard tasks, without the workbook-quality gate.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 45.0 |
| GPT-6 Astra | 45.0 |
| Claude Opus 5 | 40.0 |
| Kimi K3 | 35.0 |
| Grok 4.6 | 20.0 |