TaxCalcBench: leaderboard
TaxCalcBench TY2025 v2 headline strict-return accuracy over saved-output runs on complex federal and state tax cases with PDF inputs. Each model/tool row uses its best supported thinking-budget result; selected efforts are retained where source notes disclose them. The Gemini 3.8 Flash web row reports 40 completed cases of 50. Distinct from the older TY2024 edition and lenient or per-line metrics.
Metric: TY2025 v2 best-thinking headline Correct Returns - Strict (%). Source: github.com. Saturation forecast: Around August 2027. 47 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 40 | #3 |
| 2 | Claude Fable 5.1 | 40 | #9 |
| 3 | GPT-6 Astra | 38 | #6 |
| 4 | GPT-6.1 Sol | 36 | #7 |
| 5 | Kimi K3 | 32 | #30 |
| 6 | GPT-5.6 Sol | 30 | #19 |
| 7 | Claude Fable 5 | 28 | #16 |
| 8 | GPT-6 Sol | 28 | #14 |
| 9 | GPT-5.5 | 26 | #31 |
| 10 | Claude Opus 5 | 22 | #12 |
| 11 | Muse Wonder | 22 | #15 |
| 12 | Claude Opus 4.8 | 20 | #37 |
| 13 | Claude Sonnet 5.5 | 18 | #11 |
| 14 | GPT-6 Luna | 14 | #56 |
| 15 | Gemini 3.7 Flash | 12 | #23 |
No result here: #8 Gemini 4 Argon, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=taxcalcbench · How It Works · Data refreshed daily, snapshot 2026-10-20.