Skip to content

Benchmarks · Vals AI

Tax Agent Bench

Measures an agent on research-grade tax questions (US corporate tax) with citation scoring and category results (fact patterns, source lookup, currency, calculations, forms, controversy and precedence).

Models measured
82
Top
51.78 %Gemini 4 Carbonwithin the standard error of ±3.60
Version
1
Measured
9 November 2026
Go to the sourceThis benchmark does not count towards the PrudentScore.

01

Ranking

At the top: Top 15

Vals AI Legal Research Bench, all-pass [1]. Hatched bars are rejected by the compliance gate.

  1. 1Gemini 4 Carbon51.78%
  2. 2Claude Fable 5.150.75%
  3. 3Gemini 4 Argon49.22%
  4. 4Claude Opus 547.10%
  5. 5Claude Opus 5.544.54%
  6. 6DeepSeek V4.1 Flash 101744.35%
  7. 7GLM 5.343.33%
  8. 8Wonder Full43.00%
  9. 9Muse Spark 1.3 (Max)42.44%
  10. 10Claude Sonnet 5.541.96%
  11. 11Inkling 2 Small38.91%
  12. 12Claude Fable 538.28%
  13. 13Muse Ember38.23%
  14. 14Muse Spark 1.337.89%
  15. 15Wonder36.27%

Source: Vals AI (Tax Agent Bench) · Licence: CC BY 4.0 · Adaptation: Prudai selected the models, translated the explanations, added the gate verdict and the PrudentScore and rounded figures to two decimals.

Passes the compliance gateRejected by the compliance gate
Tax Agent Bench 1
# ModelScoreCost/taskGatePrudentScore
1Gemini 4 CarbonGoogle51.78%± 3.60$6.34Passes77
2Claude Fable 5.1Anthropic50.75%± 3.61$13.18Rejectedno score
3Gemini 4 ArgonGoogle49.22%± 3.61$3.47Passes79
4Claude Opus 5Anthropic47.10%± 3.60$5.14Passes72
5Claude Opus 5.5Anthropic44.54%± 3.58$14.40Passes63
6DeepSeek V4.1 Flash 1017DeepSeek44.35%± 3.57$0.46Rejectedno score
7GLM 5.3Z.ai43.33%± 3.56$1.80Rejectedno score
8Wonder FullMeta43.00%± 3.57$2.15Rejectedno score
9Muse Spark 1.3 (Max)Meta42.44%± 3.51$0.39Rejectedno score
10Claude Sonnet 5.5Anthropic41.96%± 3.56$8.73Passes62
11Inkling 2 SmallThinking Machines38.91%± 3.27$0.14Rejectedno score
12Claude Fable 5Anthropic38.28%± 3.50$6.70Rejectedno score
13Muse EmberMeta38.23%± 3.50$0.03Rejectedno score
14Muse Spark 1.3Meta37.89%± 3.45$0.32Rejectedno score
15WonderMeta36.27%± 3.47$1.88Rejectedno score
16Grok 4.7xAI (SpaceXAI)35.75%± 3.46$1.80Rejectedno score
17Hy4Tencent34.40%± 3.30$1.69Rejectedno score
18Grok 4.6xAI (SpaceXAI)33.66%± 3.41$0.98Rejectedno score
19Kimi K3Moonshot AI33.37%± 3.39$2.79Rejectedno score
20Muse Spark 1.2Meta32.80%± 2.75$0.27Rejectedno score
21Hy4 PreviewTencent32.62%± 3.38$0.59Rejectedno score
22MiMo V2.6 ProXiaomi32.39%± 3.36$0.13Rejectedno score
23MiniMax-M3.1-FlashMiniMax32.14%± 3.34$0.32Rejectedno score
24Gemini 3.8 FlashGoogle32.12%± 3.37$0.86Passes62
25Qwen 3.8 MaxAlibaba32.09%± 3.37$1.55Not yet assessedno score
26Claude Fable 5.2Anthropic32.09%± 3.37$3.31Rejectedno score
27Claude Haiku 5.5Anthropic31.92%± 3.35$1.16Passes59
28GLM 5.3 FlashZ.ai31.55%± 3.35$0.14Rejectedno score
29GPT-5.6 SolOpenAI31.09%± 3.34$7.51Passes57
30Ember-1Fireworks AI31.08%± 3.34$2.97Not yet assessedno score
31Qwen 3.8 27BAlibaba30.89%± 3.32$0.82Rejectedno score
32Mistral Large 4 (GA, 26.10)Mistral AI30.67%± 3.30$0.50Not yet assessedno score
33Mistral Large 4 (preview)Mistral AI30.39%± 3.31$1.39Rejectedno score
34GPT-5.6 TerraOpenAI30.04%± 3.31$4.83Passes53
35Claude Opus 4.8Anthropic30.00%± 3.30$1.69Passes62
36Claude Sonnet 5Anthropic29.37%± 3.25$1.79Passes59
37MiMo V2.6 FlashXiaomi28.51%± 3.14$0.05Rejectedno score
38DeepSeek V4 Flash 0731DeepSeek28.07%± 3.16$0.16Rejectedno score
39Step 5 PreviewStepFun27.27%± 3.11$0.55Rejectedno score
40DeepSeek V4 Pro (0813)DeepSeek26.76%± 3.09$0.73Rejectedno score
41DeepSeek V4.1 FlashDeepSeek26.63%± 3.14$0.13Rejectedno score
42Muse Spark OpenMeta26.32%± 2.65$0.88Rejectedno score
43Claude Sonnet 4.6Anthropic25.84%± 3.15$1.19Passes55
44DeepSeek V4 Pro (2026-04-23)DeepSeek25.17%± 3.06$0.72Rejectedno score
45Kimi K3.1 ProMoonshot AI24.87%± 3.12$3.64Rejectedno score
46Grok 4.5xAI (SpaceXAI)24.86%± 3.12$0.47Rejectedno score
47GPT-5.6 LunaOpenAI24.82%± 3.11$0.47Passes56
48Claude Opus 4.7Anthropic23.74%± 3.06$1.27Passes54
49Gemini 3.7 FlashGoogle22.80%± 3.03$0.48Passes58
50GPT 5.5OpenAI21.76%± 2.98$4.07Passes51
51GPT-6.1 SolOpenAI21.75%± 2.98$1.32Passes59
52Gemini 3.5 FlashGoogle21.65%± 2.96$0.86Passes47
53GPT-6 AstraOpenAI20.72%± 2.92$5.80Passes56
54MiniMax-M3MiniMax20.59%± 2.87$0.16Rejectedno score
55GPT-6 LunaOpenAI19.95%± 2.87$0.19Passes54
56Mercury 3Inception Labs18.17%± 2.54$0.05Not yet assessedno score
57Gemini 3.6 FlashGoogle18.04%± 2.77$0.36Passes46
58GPT-6.1 TerraOpenAI17.08%± 2.71$2.16Passes44
59Qwen 3.7 MaxAlibaba15.83%± 2.62$0.44Not yet assessedno score
60Inkling SmallThinking Machines15.39%± 2.37$0.09Rejectedno score
61InklingThinking Machines15.29%± 2.42$0.35Rejectedno score
62GPT-6 SolOpenAI15.00%± 2.58$2.28Passes45
63Kimi K2.6Moonshot AI12.40%± 2.38$0.57Rejectedno score
64Gemini 3 Flash (12/25)Google11.84%± 2.32$0.10Rejectedno score
65Nemotron 3.5 Super VLNVIDIA11.55%± 2.19$0.05Rejectedno score
66MiMo V2.5 ProXiaomi11.50%± 1.96$0.04Rejectedno score
67Qwen 3.7 PlusAlibaba10.47%± 1.99$0.09Not yet assessedno score
68Gemini 3.1 Pro Preview (02/26)Google9.29%± 2.09$0.30Rejectedno score
69=Beam 1.0Reflection AI8.81%± 1.90-Rejectedno score
69=GPT 5.4 MiniOpenAI8.81%± 2.04$0.81Passes25
71MiMo V2.5Xiaomi7.77%± 1.78$0.02Rejectedno score
72Qwen 3.6 PlusAlibaba7.68%± 1.81$0.30Not yet assessedno score
73Gemini 3.5 Flash LiteGoogle7.25%± 1.87$0.08Passes38
74Grok 4.20 (Reasoning)xAI (SpaceXAI)6.95%± 1.75$0.12Rejectedno score
75MiniMax-M2.7MiniMax6.87%± 1.68$0.12Rejectedno score
76Grok 4.3xAI (SpaceXAI)6.18%± 1.73$0.12Rejectedno score
77Claude Haiku 4.5 (Thinking)Anthropic5.86%± 1.54$0.23Passes32
78GPT 5.4 NanoOpenAI4.97%± 1.42$0.07Passes29
79Mistral Medium 3.5Mistral AI4.48%± 1.47$1.44Passes23
80Gemini 3.1 Flash Lite PreviewGoogle3.11%± 1.25$0.01Not yet assessedno score
81Nemotron 3.5 LightningNVIDIA2.59%± 1.15$0.03Rejectedno score
82Mercury 2.5Inception Labs2.49%± 1.02$0.02Not yet assessedno score

Source: Vals AI (Tax Agent Bench) · Licence: CC BY 4.0 · Adaptation: Prudai selected the models, translated the explanations, added the gate verdict and the PrudentScore and rounded figures to two decimals.

Vals AI · Version 1 · Measured 9 November 2026 · Go to the source. The PrudentScore is a composition by Prudai B.V. based on publicly published figures and does not represent the view or endorsement of Vals AI, Harvey or Artificial Analysis. Figures from Artificial Analysis appear in running text only, with attribution and a link, as that party's terms require [9].

Other benchmarks

Claude and Anthropic, GPT and OpenAI, Gemini and Google, Grok and xAI, Kimi and Moonshot AI, Muse and Meta, Mistral, DeepSeek, GLM and Z.ai, Harvey, Vals AI and Artificial Analysis are trademarks of their respective owners. They are named for identification only and this implies no partnership, sponsorship or endorsement.