BenchLeader
Menu
LeaderboardsModelsProvidersBenchmarksCompareReleases1BlogMethodology

Tax Agent Bench

Agentic US corporate tax research. Run by Vals AI.

As of 19 Sept 2026, Claude Fable 5.1 leads Tax Agent Bench on BenchLeader with 77.6%, ahead of Claude Opus 5 at 75.1%, across 19 model configurations with a published result.

Published by
Vals AI
Category
Agents & tools
Index weight
Reference only
Models
19
Data as of
19 Sept 2026

Vals AI (vals.ai).

What the test looks like

Corporate tax questions that require researching the code and regulations with tools, graded by tax professionals.

How it is scored

Accuracy, run by Vals AI.

What to keep in mind

US corporate tax only; rules change yearly.

Top 15
  1. 1Claude Fable 5.177.6%
  2. 2Claude Opus 575.1%
  3. 3GLM-5.3 (max)73.1%
  4. 4Muse Spark 1.3 (xhigh)71.9%
  5. 5Grok 4.6 (high)70.8%
  6. 6Kimi K3 (max)68.7%
  7. 7GPT-5.6 Sol (max)68.0%
  8. 8Gemini 3.8 Flash (high)66.8%
  9. 9Qwen3 8 (max)66.0%
  10. 10GPT-5.6 Terra (max)65.2%
  11. 11GPT-6 Astra (max)63.3%
  12. 12DeepSeek V4.1 Flash (high)62.5%
  13. 13Claude Sonnet 562.3%
  14. 14GPT-5.6 Luna (max)60.8%
  15. 15GPT-5.5 (xhigh)60.5%
19 of 19
#
1Claude Fable 5.1Anthropic77.6%64.52026-09-16
2Claude Opus 5Anthropic75.1%66.72026-09-16
3GLM-5.3maxZhipu AIopen ↗73.1%65.82026-09-16
4Muse Spark 1.3xhighMeta71.9%68.02026-09-16
5Grok 4.6highxAI70.8%64.52026-09-16
6Kimi K3maxMoonshot AIopen ↗68.7%67.22026-09-16
7GPT-5.6 SolmaxOpenAI68.0%68.82026-09-16
8Gemini 3.8 FlashhighGoogle66.8%64.52026-09-16
9Qwen3 8maxAlibaba66.0%66.52026-09-16
10GPT-5.6 TerramaxOpenAI65.2%65.02026-09-16
11GPT-6 AstramaxOpenAI63.3%71.82026-09-16
12DeepSeek V4.1 FlashhighDeepSeekopen ↗62.5%–2026-09-16
13Claude Sonnet 5Anthropic62.3%57.22026-09-16
14GPT-5.6 LunamaxOpenAI60.8%60.22026-09-16
15GPT-5.5xhighOpenAI60.5%67.62026-09-16
16DeepSeek V4 Pro 0813maxDeepSeekopen ↗58.7%56.02026-09-16
17Gemini 3.7 FlashhighGoogle57.7%64.62026-09-16
18MiniMax-M3MiniMaxopen ↗49.7%57.52026-09-16
19Mercury 2.5highInception12.8%–2026-09-16

Cite as: BenchLeader, “Tax Agent Bench leaderboard”, https://www.benchleader.com/benchmarks/vals_tax_agent, data as of 19 Sept 2026.

Tax Agent Bench: questions

What does Tax Agent Bench measure?
Corporate tax questions that require researching the code and regulations with tools, graded by tax professionals. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads Tax Agent Bench?
Claude Fable 5.1 leads Tax Agent Bench with 77.6% as of 19 Sept 2026, ahead of Claude Opus 5 at 75.1%.
How many models have Tax Agent Bench results?
19 model configurations have a Tax Agent Bench result on BenchLeader, all taken from Vals AI.
Who runs Tax Agent Bench and how often is it updated?
Tax Agent Bench is published by Vals AI. BenchLeader re-reads the published results every morning and records the date each result was published.
Does Tax Agent Bench count toward the BenchLeader Index?
No. Tax Agent Bench is shown for reference but left out of the composite index.