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
01
Ranking
At the top: Top 15
Vals AI Legal Research Bench, all-pass [1]. Hatched bars are rejected by the compliance gate.
- 1Gemini 4 Carbon51.78%
- 2Claude Fable 5.150.75%
- 3Gemini 4 Argon49.22%
- 4Claude Opus 547.10%
- 5Claude Opus 5.544.54%
- 6DeepSeek V4.1 Flash 101744.35%
- 7GLM 5.343.33%
- 8Wonder Full43.00%
- 9Muse Spark 1.3 (Max)42.44%
- 10Claude Sonnet 5.541.96%
- 11Inkling 2 Small38.91%
- 12Claude Fable 538.28%
- 13Muse Ember38.23%
- 14Muse Spark 1.337.89%
- 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.
| # Model | Score | Cost/task | Gate | PrudentScore |
|---|---|---|---|---|
| 1Gemini 4 CarbonGoogle | 51.78%± 3.60 | $6.34 | Passes | 77 |
| 2Claude Fable 5.1Anthropic | 50.75%± 3.61 | $13.18 | Rejected | no score |
| 3Gemini 4 ArgonGoogle | 49.22%± 3.61 | $3.47 | Passes | 79 |
| 4Claude Opus 5Anthropic | 47.10%± 3.60 | $5.14 | Passes | 72 |
| 5Claude Opus 5.5Anthropic | 44.54%± 3.58 | $14.40 | Passes | 63 |
| 6DeepSeek V4.1 Flash 1017DeepSeek | 44.35%± 3.57 | $0.46 | Rejected | no score |
| 7GLM 5.3Z.ai | 43.33%± 3.56 | $1.80 | Rejected | no score |
| 8Wonder FullMeta | 43.00%± 3.57 | $2.15 | Rejected | no score |
| 9Muse Spark 1.3 (Max)Meta | 42.44%± 3.51 | $0.39 | Rejected | no score |
| 10Claude Sonnet 5.5Anthropic | 41.96%± 3.56 | $8.73 | Passes | 62 |
| 11Inkling 2 SmallThinking Machines | 38.91%± 3.27 | $0.14 | Rejected | no score |
| 12Claude Fable 5Anthropic | 38.28%± 3.50 | $6.70 | Rejected | no score |
| 13Muse EmberMeta | 38.23%± 3.50 | $0.03 | Rejected | no score |
| 14Muse Spark 1.3Meta | 37.89%± 3.45 | $0.32 | Rejected | no score |
| 15WonderMeta | 36.27%± 3.47 | $1.88 | Rejected | no score |
| 16Grok 4.7xAI (SpaceXAI) | 35.75%± 3.46 | $1.80 | Rejected | no score |
| 17Hy4Tencent | 34.40%± 3.30 | $1.69 | Rejected | no score |
| 18Grok 4.6xAI (SpaceXAI) | 33.66%± 3.41 | $0.98 | Rejected | no score |
| 19Kimi K3Moonshot AI | 33.37%± 3.39 | $2.79 | Rejected | no score |
| 20Muse Spark 1.2Meta | 32.80%± 2.75 | $0.27 | Rejected | no score |
| 21Hy4 PreviewTencent | 32.62%± 3.38 | $0.59 | Rejected | no score |
| 22MiMo V2.6 ProXiaomi | 32.39%± 3.36 | $0.13 | Rejected | no score |
| 23MiniMax-M3.1-FlashMiniMax | 32.14%± 3.34 | $0.32 | Rejected | no score |
| 24Gemini 3.8 FlashGoogle | 32.12%± 3.37 | $0.86 | Passes | 62 |
| 25Qwen 3.8 MaxAlibaba | 32.09%± 3.37 | $1.55 | Not yet assessed | no score |
| 26Claude Fable 5.2Anthropic | 32.09%± 3.37 | $3.31 | Rejected | no score |
| 27Claude Haiku 5.5Anthropic | 31.92%± 3.35 | $1.16 | Passes | 59 |
| 28GLM 5.3 FlashZ.ai | 31.55%± 3.35 | $0.14 | Rejected | no score |
| 29GPT-5.6 SolOpenAI | 31.09%± 3.34 | $7.51 | Passes | 57 |
| 30Ember-1Fireworks AI | 31.08%± 3.34 | $2.97 | Not yet assessed | no score |
| 31Qwen 3.8 27BAlibaba | 30.89%± 3.32 | $0.82 | Rejected | no score |
| 32Mistral Large 4 (GA, 26.10)Mistral AI | 30.67%± 3.30 | $0.50 | Not yet assessed | no score |
| 33Mistral Large 4 (preview)Mistral AI | 30.39%± 3.31 | $1.39 | Rejected | no score |
| 34GPT-5.6 TerraOpenAI | 30.04%± 3.31 | $4.83 | Passes | 53 |
| 35Claude Opus 4.8Anthropic | 30.00%± 3.30 | $1.69 | Passes | 62 |
| 36Claude Sonnet 5Anthropic | 29.37%± 3.25 | $1.79 | Passes | 59 |
| 37MiMo V2.6 FlashXiaomi | 28.51%± 3.14 | $0.05 | Rejected | no score |
| 38DeepSeek V4 Flash 0731DeepSeek | 28.07%± 3.16 | $0.16 | Rejected | no score |
| 39Step 5 PreviewStepFun | 27.27%± 3.11 | $0.55 | Rejected | no score |
| 40DeepSeek V4 Pro (0813)DeepSeek | 26.76%± 3.09 | $0.73 | Rejected | no score |
| 41DeepSeek V4.1 FlashDeepSeek | 26.63%± 3.14 | $0.13 | Rejected | no score |
| 42Muse Spark OpenMeta | 26.32%± 2.65 | $0.88 | Rejected | no score |
| 43Claude Sonnet 4.6Anthropic | 25.84%± 3.15 | $1.19 | Passes | 55 |
| 44DeepSeek V4 Pro (2026-04-23)DeepSeek | 25.17%± 3.06 | $0.72 | Rejected | no score |
| 45Kimi K3.1 ProMoonshot AI | 24.87%± 3.12 | $3.64 | Rejected | no score |
| 46Grok 4.5xAI (SpaceXAI) | 24.86%± 3.12 | $0.47 | Rejected | no score |
| 47GPT-5.6 LunaOpenAI | 24.82%± 3.11 | $0.47 | Passes | 56 |
| 48Claude Opus 4.7Anthropic | 23.74%± 3.06 | $1.27 | Passes | 54 |
| 49Gemini 3.7 FlashGoogle | 22.80%± 3.03 | $0.48 | Passes | 58 |
| 50GPT 5.5OpenAI | 21.76%± 2.98 | $4.07 | Passes | 51 |
| 51GPT-6.1 SolOpenAI | 21.75%± 2.98 | $1.32 | Passes | 59 |
| 52Gemini 3.5 FlashGoogle | 21.65%± 2.96 | $0.86 | Passes | 47 |
| 53GPT-6 AstraOpenAI | 20.72%± 2.92 | $5.80 | Passes | 56 |
| 54MiniMax-M3MiniMax | 20.59%± 2.87 | $0.16 | Rejected | no score |
| 55GPT-6 LunaOpenAI | 19.95%± 2.87 | $0.19 | Passes | 54 |
| 56Mercury 3Inception Labs | 18.17%± 2.54 | $0.05 | Not yet assessed | no score |
| 57Gemini 3.6 FlashGoogle | 18.04%± 2.77 | $0.36 | Passes | 46 |
| 58GPT-6.1 TerraOpenAI | 17.08%± 2.71 | $2.16 | Passes | 44 |
| 59Qwen 3.7 MaxAlibaba | 15.83%± 2.62 | $0.44 | Not yet assessed | no score |
| 60Inkling SmallThinking Machines | 15.39%± 2.37 | $0.09 | Rejected | no score |
| 61InklingThinking Machines | 15.29%± 2.42 | $0.35 | Rejected | no score |
| 62GPT-6 SolOpenAI | 15.00%± 2.58 | $2.28 | Passes | 45 |
| 63Kimi K2.6Moonshot AI | 12.40%± 2.38 | $0.57 | Rejected | no score |
| 64Gemini 3 Flash (12/25)Google | 11.84%± 2.32 | $0.10 | Rejected | no score |
| 65Nemotron 3.5 Super VLNVIDIA | 11.55%± 2.19 | $0.05 | Rejected | no score |
| 66MiMo V2.5 ProXiaomi | 11.50%± 1.96 | $0.04 | Rejected | no score |
| 67Qwen 3.7 PlusAlibaba | 10.47%± 1.99 | $0.09 | Not yet assessed | no score |
| 68Gemini 3.1 Pro Preview (02/26)Google | 9.29%± 2.09 | $0.30 | Rejected | no score |
| 69=Beam 1.0Reflection AI | 8.81%± 1.90 | - | Rejected | no score |
| 69=GPT 5.4 MiniOpenAI | 8.81%± 2.04 | $0.81 | Passes | 25 |
| 71MiMo V2.5Xiaomi | 7.77%± 1.78 | $0.02 | Rejected | no score |
| 72Qwen 3.6 PlusAlibaba | 7.68%± 1.81 | $0.30 | Not yet assessed | no score |
| 73Gemini 3.5 Flash LiteGoogle | 7.25%± 1.87 | $0.08 | Passes | 38 |
| 74Grok 4.20 (Reasoning)xAI (SpaceXAI) | 6.95%± 1.75 | $0.12 | Rejected | no score |
| 75MiniMax-M2.7MiniMax | 6.87%± 1.68 | $0.12 | Rejected | no score |
| 76Grok 4.3xAI (SpaceXAI) | 6.18%± 1.73 | $0.12 | Rejected | no score |
| 77Claude Haiku 4.5 (Thinking)Anthropic | 5.86%± 1.54 | $0.23 | Passes | 32 |
| 78GPT 5.4 NanoOpenAI | 4.97%± 1.42 | $0.07 | Passes | 29 |
| 79Mistral Medium 3.5Mistral AI | 4.48%± 1.47 | $1.44 | Passes | 23 |
| 80Gemini 3.1 Flash Lite PreviewGoogle | 3.11%± 1.25 | $0.01 | Not yet assessed | no score |
| 81Nemotron 3.5 LightningNVIDIA | 2.59%± 1.15 | $0.03 | Rejected | no score |
| 82Mercury 2.5Inception Labs | 2.49%± 1.02 | $0.02 | Not yet assessed | no 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
- Legal Research Bench
- Harvey LAB (Vals implementation)
- LegalBench (Vals leaderboard)
- CaseLaw v2
- TaxEval v2
- CorpFin v2
- Immigration Bench
- Access to Justice
- Public Benefits Bench
- Vals Legal Index
- Vals Index
- GPQA Diamond (Vals-run)
- MMLU Pro (Vals-run)
- Artificial Analysis Intelligence Index
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.