Skip to content

Benchmarks · Vals AI

Harvey LAB (Vals implementation)

Note: Vals has no benchmark aimed at Dutch or EU law. Measures whether an agent performs real legal work with documents, spreadsheets, presentations and file-system tools (tasks from Harvey's Legal Agent Benchmark, 24 practice areas). "All-pass" counts only when every rubric criterion of a task is met; "criteria-pass-rate" is the share of criteria met. Version 1.1 has been the default since 19 October 2026; version 1 is archived (last update 18 October 2026) and is what the current PrudentBench page cites.

Models measured
80
Top
25.42 %Muse Spark 1.2within the standard error of ±3.67
Version
1
Measured
18 October 2026
Go to the sourceIn the PrudentScore, version 1 of this benchmark counts for 15 percent.
Versions

01

Ranking

At the top: Top 15

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

  1. 1Muse Spark 1.225.42%
  2. 2Muse Spark 1.3 (Max)23.75%
  3. 3Muse Spark 1.322.92%
  4. 4Wonder Full21.67%
  5. 5Muse Spark 1.120.00%
  6. 6Gemini 4 Argon19.58%
  7. 7GLM Legal19.17%
  8. 8Wonder17.08%
  9. 9Muse Spark Open16.67%
  10. 10=Grok 4.615.83%
  11. 10=Mistral Large 4 (preview)15.83%
  12. 12=Grok 4.512.92%
  13. 12=Kimi K312.92%
  14. 14Grok 4.712.50%
  15. 15=Claude Fable 511.25%

Source: Vals AI (Harvey LAB (Vals implementation)) · 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
Harvey LAB (Vals implementation) 1
# ModelScoreCost/taskGatePrudentScore
1Muse Spark 1.2Meta25.42%± 3.67$2.09Rejectedno score
2Muse Spark 1.3 (Max)Meta23.75%± 3.55$2.26Rejectedno score
3Muse Spark 1.3Meta22.92%± 3.42$2.76Rejectedno score
4Wonder FullMeta21.67%± 3.54$5.41Rejectedno score
5Muse Spark 1.1Meta20.00%± 2.94$0.80Rejectedno score
6Gemini 4 ArgonGoogle19.58%± 3.31$10.55Passes79
7GLM LegalZ.ai19.17%± 3.25$0.15Rejectedno score
8WonderMeta17.08%± 3.10$7.08Rejectedno score
9Muse Spark OpenMeta16.67%± 3.09$6.83Rejectedno score
10=Grok 4.6xAI (SpaceXAI)15.83%± 2.65$4.01Rejectedno score
10=Mistral Large 4 (preview)Mistral AI15.83%± 2.96$4.18Rejectedno score
12=Grok 4.5xAI (SpaceXAI)12.92%± 2.53$2.02Rejectedno score
12=Kimi K3Moonshot AI12.92%± 2.68$3.83Rejectedno score
14Grok 4.7xAI (SpaceXAI)12.50%± 2.66$11.13Rejectedno score
15=Claude Fable 5Anthropic11.25%± 2.15$19.23Rejectedno score
15=MiMo V2.6 FlashXiaomi11.25%± 2.54$0.09Rejectedno score
15=Qwen 3.8 27BAlibaba11.25%± 2.42$4.03Rejectedno score
18MiMo V2.6 ProXiaomi10.83%± 2.45$0.22Rejectedno score
19Qwen 3.8 MaxAlibaba10.42%± 2.29$2.37Not yet assessedno score
20Gemini 3.8 FlashGoogle10.00%± 2.42$3.66Passes62
21Claude Opus 4.8Anthropic9.58%± 2.15$10.22Passes62
22=DeepSeek V4.1 Flash 1017DeepSeek9.17%± 2.51$0.54Rejectedno score
22=Ember-1Fireworks AI9.17%± 2.44$3.33Not yet assessedno score
22=Hy4 PreviewTencent9.17%± 2.00$0.98Rejectedno score
22=Step 5 PreviewStepFun9.17%± 2.44$1.00Rejectedno score
26Gemini 3.7 FlashGoogle8.75%± 2.15$2.56Passes58
27=DeepSeek V4 Flash 0731DeepSeek8.33%± 2.15$0.06Rejectedno score
27=GLM 5.3Z.ai8.33%± 2.00$4.30Rejectedno score
29DeepSeek V4 Pro (0813)DeepSeek7.50%± 2.15$0.17Rejectedno score
30GLM 5.2Z.ai7.08%± 2.00$2.06Rejectedno score
31=Claude Fable 5.1Anthropic6.67%± 1.17$46.21Rejectedno score
31=Claude Opus 4.7Anthropic6.67%± 2.00$10.81Passes54
31=Claude Opus 5Anthropic6.67%± 1.65$23.67Passes72
31=DeepSeek V4.1 FlashDeepSeek6.67%± 1.83$0.17Rejectedno score
31=GLM 5.3 FlashZ.ai6.67%± 1.65$0.57Rejectedno score
36=GPT-6 AstraOpenAI5.42%± 1.17$26.16Passes56
36=GPT-6.1 SolOpenAI5.42%± 1.75$3.93Passes59
38=Claude Sonnet 4.6Anthropic5.00%± 1.65$3.04Passes55
38=Claude Sonnet 5Anthropic5.00%± 1.65$8.95Passes59
40MiniMax-M3MiniMax4.17%± 1.65$1.46Rejectedno score
41=Claude Opus 5.5Anthropic3.75%± 1.47$21.38Passes63
41=DeepSeek V4 Pro (2026-04-23)DeepSeek3.75%± 1.43$0.74Rejectedno score
41=GPT 5.5OpenAI3.75%± 1.17$4.60Passes51
44Gemini 3.6 FlashGoogle3.33%± 1.17$1.90Passes46
45=Claude Sonnet 5.5Anthropic2.92%± 1.36$16.61Passes62
45=GPT-6 LunaOpenAI2.92%± 1.36$0.30Passes54
47=Gemini 3.5 FlashGoogle2.50%± 0.83$2.12Passes47
47=GPT-5.6 SolOpenAI2.50%± 0.83$10.37Passes57
49=InklingThinking Machines2.08%± 0.83$1.25Rejectedno score
49=MiMo V2.5 ProXiaomi2.08%± 0.83$0.11Rejectedno score
51=GPT-6 SolOpenAI1.67%± 0.82$3.36Passes45
51=Inkling SmallThinking Machines1.67%± 0.83$0.17Rejectedno score
51=Kimi K2.6Moonshot AI1.67%± 0.83$0.82Rejectedno score
51=MiMo V2.5Xiaomi1.67%± 0.00$0.04Rejectedno score
51=Qwen 3.7 MaxAlibaba1.67%± 0.83$1.58Not yet assessedno score
56=Claude Haiku 5.5Anthropic1.25%± 0.72$3.39Passes59
56=GPT-5.6 LunaOpenAI1.25%± 0.83$0.38Passes56
56=Ling 3.0 FlashAnt Group1.25%± 0.83$0.05Rejectedno score
56=Qwen 3.6 PlusAlibaba1.25%± 0.83$3.94Not yet assessedno score
60=Claude Haiku 4.5 (Thinking)Anthropic0.83%± 0.83$0.42Passes32
60=GPT-5.6 TerraOpenAI0.83%± 0.83$3.20Passes53
62=Grok 4.3xAI (SpaceXAI)0.42%± 0.00$0.39Rejectedno score
62=Mistral Medium 3.5Mistral AI0.42%± 0.00$6.07Passes23
62=Nemotron 3 UltraNVIDIA0.42%± 0.00-Rejectedno score
65=Gemini 3 Flash (12/25)Google0.00%± 0.00$0.26Rejectedno score
65=Gemini 3.1 Flash Lite PreviewGoogle0.00%± 0.00$0.08Not yet assessedno score
65=Gemini 3.1 Pro Preview (02/26)Google0.00%± 0.00$1.15Rejectedno score
65=Gemini 3.5 Flash LiteGoogle0.00%± 0.00$0.33Passes38
65=GLM 5.1Z.ai0.00%± 0.00$0.56Rejectedno score
65=GPT 5.4 (xhigh)OpenAI0.00%± 0.00$3.60Passes0
65=GPT 5.4 MiniOpenAI0.00%± 0.00$1.00Passes25
65=GPT 5.4 NanoOpenAI0.00%± 0.00$0.18Passes29
65=Grok 4.20 (Reasoning)xAI (SpaceXAI)0.00%± 0.00$0.99Rejectedno score
65=Kimi K2.5Moonshot AI0.00%± 0.00$0.35Passes37
65=Laguna M.1Poolside0.00%± 0.00$0.53Rejectedno score
65=Laguna XS.2Poolside0.00%± 0.00$0.28Rejectedno score
65=Mercury 2.5Inception Labs0.00%± 0.00$0.19Not yet assessedno score
65=MiniMax-M2.7MiniMax0.00%± 0.00$0.49Rejectedno score
65=Nemotron 3.5 LightningNVIDIA0.00%± 0.00$0.03Rejectedno score
65=Qwen 3.7 PlusAlibaba0.00%± 0.00$0.23Not yet assessedno score

Source: Vals AI (Harvey LAB (Vals implementation)) · 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 18 October 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.