TreasuryQA Bench pass^3
Professional Work · 2026-10-19
Share of TreasuryQA Bench's 70 exact-figure questions over scanned U.S. Treasury volumes that an agent answers correctly on all three of its attempts, from the same runs as the TreasuryQA Bench pass@1.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 45.7 |
| Gemini 4 Argon | 45.7 |
| GPT-6.1 Sol | 38.6 |
| Gemini 3.8 Flash | 37.1 |
| Claude Sonnet 5.5 | 27.1 |
| DeepSeek V4.1 Flash | 17.1 |
| GPT-5.6 Terra | 8.6 |
| Grok 4.7 | 4.3 |