Vals AI Tax Agent Bench: leaderboard
Research-grade US corporate tax questions run as agent tasks: the model has to work the question the way a tax professional would rather than answer it from memory, and Vals grades it on weighted partial credit rather than exact match. Eighteen models at launch, added to the catalogue on 2026-09-05.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around 2028. 73 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (Max) | 49.19 | #9 (Claude Fable 5.1) |
| 2 | Claude Opus 5.5 (Max) | 45.57 | #4 (Claude Opus 5.5) |
| 3 | Claude Opus 5 (Max) | 45.55 | #12 (Claude Opus 5) |
| 4 | Gemini 4 Argon (High) | 44.56 | #8 (Gemini 4 Argon) |
| 5 | DeepSeek V4.1 Flash (1017) (Max) | 44.35 | #20 (DeepSeek V4.1 Flash (1017)) |
| 6 | Muse Wonder Full (Max) | 43 | #10 (Muse Wonder Full) |
| 7 | Muse Spark 1.3 (Max) | 42.44 | #27 (Muse Spark 1.3) |
| 8 | Claude Sonnet 5.5 (Max) | 41.96 | #11 (Claude Sonnet 5.5) |
| 9 | GLM-5.3 (Max) | 39.7 | #36 (GLM-5.3) |
| 10 | Claude Fable 5 (Max) | 38.28 | #16 (Claude Fable 5) |
| 11 | Muse Wonder (Max) | 36.27 | #14 (Muse Wonder) |
| 12 | Muse Spark 1.3 (xHigh) | 35.97 | #27 (Muse Spark 1.3) |
| 13 | Grok 4.7 (xHigh) | 34.2 | #26 (Grok 4.7) |
| 14 | Kimi K3 (Max) | 33.37 | #30 (Kimi K3) |
| 15 | Grok 4.6 (High) | 33.14 | #25 (Grok 4.6) |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-tax-agent-bench · How It Works · Data refreshed daily, snapshot 2026-10-23.