TaxBench Data Retrieval pass^5
Professional Work · 2026-05-04
Data Retrieval subset of TaxBench reporting pass^5 repeated-attempt reliability performance on professional tax tasks.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 Pro | 22.9 |
| GPT-5.5 | 20.9 |
| Opus 4.6 | 19.2 |
| GPT-5.4 Pro | 13.5 |
| Gemini 3.1 Pro Preview | 12.5 |
| Opus 4.7 | 11.1 |
| GPT-5.2 Pro | 7.1 |
| Sonnet 4.6 | 7.1 |
| Sonnet 4.5 | 6.8 |
| GPT-5.4 | 6.0 |
| Grok 4.20 | 4.0 |
| Grok 4.1 Fast | 3.7 |
| GPT-5.2 | 2.7 |
| Gemini 3.1 Flash (benchmark label) | 1.4 |
| Gemini 2.5 Pro | 1.3 |