Tax Agent Bench
Whether an agent can research a research-grade US tax question and produce a supported answer — the multi-step process a corporate tax practitioner follows from authority to conclusion.
How it is run
Each task is a written question about federal corporate income tax (roughly 60% of the set), partnership tax, state and local tax, transfer pricing or corporate employment tax. The agent researches with tools for as long as it needs and submits one answer, which is graded against an expert rubric with a check that its cited sources exist; the trajectory is recorded but not scored.
What it cannot tell you
The overall score can hide differences between question categories. Forms & Filings and Controversy & Precedence Analysis require several supported elements, and a model that does well on rule lookup may still miss those requirements.
- Leader
- Claude Fable 5.177.64% · Vals AI's own list
- Top-ten spread
- 7.82%Leader minus tenth place, or last place in smaller fields.
- Median
- 59.97%of the 64 rows published
- No published result
- 9of 46 models on this site
Independent leaderboard · Vals AI
Top 64 of 64 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5.1 | 77.64% |
| 2 | Gemini 4 Argon | 76.23% |
| 3 | Claude Opus 5 | 75.06% |
| 4 | Claude Sonnet 5.5 | 73.39% |
| 5 | GLM 5.3 | 73.09% |
| 6 | Muse Spark 1.3 Max | 72.44% |
| 7 | Muse Spark 1 3 | 71.93% |
| 8 | Grok 4.6 | 70.79% |
| 9 | Claude Opus 5.5 | 70.50% |
| 10 | Claude Fable 5 | 69.82% |
| 11 | Kimi K3 | 68.67% |
| 12 | Ember 1 | 68.04% |
| 13 | GPT-5.6 Sol | 67.95% |
| 14 | Gemini 3.8 Flash | 66.77% |
| 15 | Qwen3.8 Max | 65.96% |
| 16 | Grok 4.7 | 65.60% |
| 17 | GPT-5.6 Terra | 65.20% |
| 18 | MiMo V2.6 Pro | 64.94% |
| 19 | Claude Opus 4.8 | 64.91% |
| 20 | Claude Sonnet 4 6 | 64.85% |
| 21 | Grok 4.5 | 64.32% |
| 22 | Claude Opus 4.7 | 63.86% |
| 23 | Hunyuan Hy4 Preview | 63.71% |
| 24 | GPT-6 Astra | 63.34% |
| 25 | Deepseek V4 Flash 0731 | 62.56% |
| 26 | DeepSeek V4.1 Flash | 62.46% |
| 27 | GPT-6.1 Sol | 62.31% |
| 28 | Claude Sonnet 5 | 62.27% |
| 29 | GLM 5.3 Flash | 61.96% |
| 30 | GPT-5.6 Luna | 60.81% |
| 31 | GPT-5.5 | 60.46% |
| 32 | Qwen3.8 27b | 60.04% |
| 33 | MiMo-V2.6-Flash | 59.90% |
| 34 | GPT-6 Luna | 58.86% |
| 35 | DeepSeek V4 Pro | 58.66% |
| 36 | Deepseek V4 Pro | 58.50% |
| 37 | Gemini 3.7 Flash | 57.66% |
| 38 | Muse Spark 1.2 | 56.86% |
| 39 | Gemini 3.5 Flash | 53.53% |
| 40 | GPT-6 Sol | 53.05% |
| 41 | Gemini 3.6 Flash | 49.90% |
| 42 | MiniMax M3 | 49.69% |
| 43 | Qwen3.7 Max | 47.41% |
| 44 | Ling 3.0 Flash Af Rc3 | 45.70% |
| 45 | Kimi K2.6 | 45.13% |
| 46 | Inkling | 43.52% |
| 47 | Inkling Small | 41.46% |
| 48 | Qwen3.7 Plus | 38.71% |
| 49 | Gemini 3.1 Pro | 37.84% |
| 50 | Gemini 3 Flash Preview | 37.56% |
| 51 | Gemini 3.5 Flash Lite | 37.45% |
| 52 | Gpt 5.4 Mini 2026 03 17 | 36.59% |
| 53 | Mimo V2.5 Pro | 34.79% |
| 54 | Grok 4.3 | 34.42% |
| 55 | Qwen3.6 Plus | 32.84% |
| 56 | Grok 4.20 0309 Reasoning | 31.24% |
| 57 | Mimo V2.5 | 29.16% |
| 58 | Claude Haiku 4 5 20251001 Thinking | 28.31% |
| 59 | Mistral Medium 3.5 | 27.29% |
| 60 | Gpt 5.4 Nano 2026 03 17 | 26.58% |
| 61 | MiniMax M2.7 | 25.38% |
| 62 | Gemini 3.1 Flash Lite Preview | 20.09% |
| 63 | Nemotron Lightning 3p5 30b A3b | 13.00% |
| 64 | Mercury 2.5 | 12.41% |
No published result tracked
9 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Read this with
Benchmarks in the same category, which usually disagree with this one on purpose.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.vals.ai/benchmarks/tax_agent_bench
- Independent resultswww.vals.ai/benchmarks/tax_agent_bench