Benchmarks · Vals AI
TaxEval v2
Measures answering tax questions (questions and answers created by Vals). Vals marks the leaderboard "Archived Benchmark": "Since performance on this benchmark has saturated, we no longer run this benchmark on new model releases" (last update 1 September 2026).
- Models measured
- 108
- Top
- 80.38 %Muse Spark 1.2within the standard error of ±0.76
- Version
- 2
- Measured
- 1 September 2026
01
Ranking
At the top: Top 15
Vals AI Legal Research Bench, all-pass [1]. Hatched bars are rejected by the compliance gate.
- 1Muse Spark 1.280.38%
- 2Muse Spark 1.179.72%
- 3Muse Spark77.68%
- 4Claude Sonnet 4.677.11%
- 5Claude Fable 576.94%
- 6=GPT-5.6 Luna76.17%
- 6=GPT-5.6 Terra76.17%
- 8=Claude Fable 5.175.96%
- 8=Claude Opus 4.6 (Thinking)75.96%
- 10GPT 5.275.76%
- 11Kimi K375.72%
- 12Grok 4 Fast (Reasoning)75.70%
- 13=Claude Opus 4.875.63%
- 13=Claude Sonnet 575.63%
- 15GLM 5.3 Flash75.59%
Source: Vals AI (TaxEval v2) · 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 |
|---|---|---|---|---|
| 1Muse Spark 1.2Meta | 80.38%± 0.76 | - | Rejected | no score |
| 2Muse Spark 1.1Meta | 79.72%± 0.77 | - | Rejected | no score |
| 3Muse SparkMeta | 77.68%± 0.81 | - | Rejected | no score |
| 4Claude Sonnet 4.6Anthropic | 77.11%± 0.82 | - | Passes | 55 |
| 5Claude Fable 5Anthropic | 76.94%± 0.82 | - | Rejected | no score |
| 6=GPT-5.6 LunaOpenAI | 76.17%± 0.84 | - | Passes | 56 |
| 6=GPT-5.6 TerraOpenAI | 76.17%± 0.84 | - | Passes | 53 |
| 8=Claude Fable 5.1Anthropic | 75.96%± 0.83 | - | Rejected | no score |
| 8=Claude Opus 4.6 (Thinking)Anthropic | 75.96%± 0.83 | - | Passes | no score |
| 10GPT 5.2OpenAI | 75.76%± 0.84 | - | Passes | no score |
| 11Kimi K3Moonshot AI | 75.72%± 0.84 | - | Rejected | no score |
| 12Grok 4 Fast (Reasoning)xAI (SpaceXAI) | 75.70%± 0.76 | - | Rejected | no score |
| 13=Claude Opus 4.8Anthropic | 75.63%± 0.84 | - | Passes | 62 |
| 13=Claude Sonnet 5Anthropic | 75.63%± 0.84 | - | Passes | 59 |
| 15GLM 5.3 FlashZ.ai | 75.59%± 0.84 | - | Rejected | no score |
| 16Qwen 3.8 MaxAlibaba | 75.55%± 0.84 | - | Not yet assessed | no score |
| 17Inkling SmallThinking Machines | 75.51%± 0.84 | - | Rejected | no score |
| 18=InklingThinking Machines | 75.31%± 0.84 | - | Rejected | no score |
| 18=Qwen 3.7 MaxAlibaba | 75.31%± 0.84 | - | Not yet assessed | no score |
| 20Claude Opus 4.7Anthropic | 75.27%± 0.84 | - | Passes | 54 |
| 21GPT 5 MiniOpenAI | 75.22%± 0.85 | - | Passes | no score |
| 22Claude Opus 5Anthropic | 75.14%± 0.83 | - | Passes | 72 |
| 23GPT 4.1OpenAI | 75.06%± 0.84 | - | Passes | no score |
| 24GPT 5.5OpenAI | 74.98%± 0.85 | - | Passes | 51 |
| 25=Gemini 3.6 FlashGoogle | 74.86%± 0.85 | - | Passes | 46 |
| 25=GPT 5.1OpenAI | 74.86%± 0.85 | - | Passes | no score |
| 27Claude Opus 4.5 (Thinking)Anthropic | 74.86%± 0.85 | - | Passes | no score |
| 28o4 MiniOpenAI | 74.78%± 0.85 | - | Passes | no score |
| 29GPT-5.6 SolOpenAI | 74.78%± 0.86 | - | Passes | 57 |
| 30=Gemini 3.7 FlashGoogle | 74.73%± 0.85 | - | Passes | 58 |
| 30=Qwen 3.6 PlusAlibaba | 74.73%± 0.85 | - | Not yet assessed | no score |
| 32Kimi K2.6Moonshot AI | 74.65%± 0.85 | - | Rejected | no score |
| 33o3OpenAI | 74.57%± 0.85 | - | Passes | no score |
| 34GPT 4o (2024-11-20)OpenAI | 74.53%± 0.86 | - | Passes | no score |
| 35Gemini 3.8 FlashGoogle | 74.45%± 0.85 | - | Passes | 62 |
| 36Gemini 3.5 FlashGoogle | 74.37%± 0.85 | - | Passes | 47 |
| 37Claude Opus 4.5 (Nonthinking)Anthropic | 74.33%± 0.85 | - | Passes | no score |
| 38Kimi K2.5Moonshot AI | 74.20%± 0.85 | - | Passes | 37 |
| 39Grok 4.20 (Reasoning)xAI (SpaceXAI) | 74.12%± 0.86 | - | Rejected | no score |
| 40=GPT 5.4 (xhigh)OpenAI | 73.96%± 0.87 | - | Passes | 0 |
| 40=Qwen 3 Max PreviewAlibaba | 73.96%± 0.86 | - | Not yet assessed | no score |
| 42Gemini 3 Flash (12/25)Google | 73.88%± 0.86 | - | Rejected | no score |
| 43MiMo V2.5 ProXiaomi | 73.79%± 0.86 | - | Rejected | no score |
| 44Claude Opus 4.1 (Thinking)Anthropic | 73.67%± 0.86 | - | Rejected | no score |
| 45Qwen 3 MaxAlibaba | 73.51%± 0.86 | - | Not yet assessed | no score |
| 46GPT 5OpenAI | 73.39%± 0.87 | - | Passes | no score |
| 47GLM 5.2Z.ai | 73.34%± 0.86 | - | Rejected | no score |
| 48Claude Sonnet 4.5 (Thinking)Anthropic | 73.30%± 0.86 | - | Passes | no score |
| 49Grok 4.1 Fast (Reasoning)xAI (SpaceXAI) | 73.14%± 0.86 | - | Rejected | no score |
| 50Nemotron 3 UltraNVIDIA | 73.10%± 0.87 | - | Rejected | no score |
| 51=DeepSeek V4 Pro (0813)DeepSeek | 73.06%± 0.87 | - | Rejected | no score |
| 51=Mistral Large 3Mistral AI | 73.06%± 0.87 | - | Passes | no score |
| 53Gemini 3.1 Pro Preview (02/26)Google | 72.88%± 0.86 | - | Rejected | no score |
| 54=Gemini 2.5 Flash Preview (9/25) (Nonthinking)Google | 72.73%± 0.87 | - | Rejected | no score |
| 54=MiniMax-M3MiniMax | 72.73%± 0.86 | - | Rejected | no score |
| 56Gemini 3.5 Flash LiteGoogle | 72.61%± 0.88 | - | Passes | 38 |
| 57Gemini 3 Pro (11/25)Google | 72.57%± 0.87 | - | Rejected | no score |
| 58=Gemini 2.5 Flash Preview (9/25) (Thinking)Google | 72.40%± 0.87 | - | Rejected | no score |
| 58=GLM 4.5Z.ai | 72.40%± 0.87 | - | Rejected | no score |
| 60GLM 5.3Z.ai | 72.36%± 0.88 | - | Rejected | no score |
| 61Qwen 3.5 FlashAlibaba | 72.16%± 0.88 | - | Not yet assessed | no score |
| 62DeepSeek V4 Pro (2026-04-23)DeepSeek | 72.08%± 0.88 | - | Rejected | no score |
| 63Claude Sonnet 4 (Thinking)Anthropic | 72.00%± 0.88 | - | Rejected | no score |
| 64=Claude Opus 4 (Nonthinking)Anthropic | 71.91%± 0.88 | - | Rejected | no score |
| 64=GPT 4.1 MiniOpenAI | 71.91%± 0.88 | - | Passes | no score |
| 66MiMo V2.5Xiaomi | 71.83%± 0.88 | - | Rejected | no score |
| 67Gemini 3.1 Flash Lite PreviewGoogle | 71.79%± 0.88 | - | Not yet assessed | no score |
| 68Kimi K2 ThinkingMoonshot AI | 71.71%± 0.88 | - | Rejected | no score |
| 69Grok 4.5xAI (SpaceXAI) | 71.67%± 0.89 | - | Rejected | no score |
| 70GPT OSS 120BOpenAI | 71.59%± 0.88 | - | Passes | no score |
| 71Grok 4 Fast (Non-Reasoning)xAI (SpaceXAI) | 71.58%± 0.78 | - | Rejected | no score |
| 72Claude Opus 4.1 (Nonthinking)Anthropic | 71.46%± 0.88 | - | Rejected | no score |
| 73Qwen 3.6 27BAlibaba | 71.26%± 0.88 | - | Rejected | no score |
| 74GPT 5.4 MiniOpenAI | 71.22%± 0.90 | - | Passes | 25 |
| 75GLM 5.1Z.ai | 71.19%± 0.90 | - | Rejected | no score |
| 76Grok 4.6xAI (SpaceXAI) | 71.10%± 0.90 | - | Rejected | no score |
| 77Qwen 3.8 27BAlibaba | 70.85%± 0.89 | - | Rejected | no score |
| 78Grok 4.3xAI (SpaceXAI) | 70.81%± 0.90 | - | Rejected | no score |
| 79DeepSeek V4 Flash 0731DeepSeek | 70.69%± 0.90 | - | Rejected | no score |
| 80Ling 3.0 FlashAnt Group | 70.65%± 0.90 | - | Rejected | no score |
| 81Mistral Medium 3.1 (05/2025)Mistral AI | 70.32%± 0.89 | - | Passes | no score |
| 82Kimi K2 InstructMoonshot AI | 70.20%± 0.90 | - | Rejected | no score |
| 83GLM 5Z.ai | 70.03%± 0.90 | - | Passes | no score |
| 84Claude Sonnet 4 (Nonthinking)Anthropic | 69.62%± 0.90 | - | Rejected | no score |
| 85GLM 4.7Z.ai | 68.77%± 0.92 | - | Passes | no score |
| 86=DeepSeek V3.2 (Thinking)DeepSeek | 68.15%± 0.92 | - | Passes | no score |
| 86=MiniMax-M2.5MiniMax | 68.15%± 0.91 | - | Passes | no score |
| 88Mistral Medium 3.5Mistral AI | 67.99%± 0.92 | - | Passes | 23 |
| 89Claude Haiku 4.5 (Thinking)Anthropic | 67.54%± 0.91 | - | Passes | 32 |
| 90GPT 5.4 NanoOpenAI | 67.42%± 0.93 | - | Passes | 29 |
| 91GPT 5 NanoOpenAI | 67.38%± 0.91 | - | Passes | no score |
| 92=Llama 4 MaverickMeta | 66.56%± 0.92 | - | Rejected | no score |
| 92=MiniMax-M2.7MiniMax | 66.56%± 0.92 | - | Rejected | no score |
| 94MiniMax-M2.1MiniMax | 66.35%± 0.93 | - | Passes | no score |
| 95GLM 4.6Z.ai | 66.23%± 0.81 | - | Rejected | no score |
| 96Gemini 2.5 Flash Lite (9/25) (Nonthinking)Google | 66.23%± 0.92 | - | Rejected | no score |
| 97Grok 4xAI (SpaceXAI) | 65.09%± 0.96 | - | Rejected | no score |
| 98Gemini 2.5 Flash Lite (9/25) (Thinking)Google | 64.72%± 0.93 | - | Rejected | no score |
| 99GPT OSS 20BOpenAI | 63.70%± 0.94 | - | Passes | no score |
| 100Magistral Medium 1.2 (09/2025)Mistral AI | 61.94%± 0.96 | - | Not yet assessed | no score |
| 101Grok 4.1 Fast Non-ReasoningxAI (SpaceXAI) | 61.86%± 0.91 | - | Rejected | no score |
| 102Command ACohere | 61.37%± 0.95 | - | Rejected | no score |
| 103GPT 4.1 NanoOpenAI | 60.75%± 0.96 | - | Passes | no score |
| 104Magistral Small 1.2 (09/2025)Mistral AI | 60.30%± 0.97 | - | Passes | no score |
| 105Laguna XS.2Poolside | 58.95%± 0.96 | - | Rejected | no score |
| 106Mistral Small 3.1 (03/2025)Mistral AI | 58.30%± 0.97 | - | Passes | no score |
| 107Llama 4 ScoutMeta | 55.19%± 0.97 | - | Rejected | no score |
| 108Laguna M.1Poolside | 1.64%± 0.26 | - | Rejected | no score |
Source: Vals AI (TaxEval v2) · 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 2 · Measured 1 September 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
- Tax Agent Bench
- 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.