SAB Research Withheld Facts Answered
Finance & Accounting
Share of the 221 withheld criteria in SAB's research tasks, spread over 95 tasks, for which a model gave a value instead of unknown after the benchmark removed the documents that support them. Lower is better; a value on any withheld criterion also fails its task. It comes from the same runs and v0.7.1 answer keys as the research score.
Top models (lower is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 8.6 |
| GPT-6 Astra | 10.4 |
| GPT-6 Sol | 14.9 |
| Claude Opus 5 | 17.6 |
| GPT-6 Luna | 19.5 |
| Claude Fable 5.1 | 19.9 |
| Haiku 4.5 | 21.3 |
| DeepSeek-V3.2 | 21.3 |
| Sonnet 5 | 29.9 |
| Grok 4.6 | 29.9 |
| GPT-5.6 Luna | 30.8 |
| GPT-5.6 Sol | 32.1 |
| GPT-5.6 Terra | 32.6 |