Skip to content

Benchmarks · Vals AI

Public Benefits Bench

Measures whether AI can help people navigate US SNAP benefits: questions on benefits workflows.

Models measured
50
Top
76.93 %Claude Opus 5within the standard error of ±1.10
Version
1.1
Measured
12 October 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. 1Claude Opus 576.93%
  2. 2Claude Fable 5.174.90%
  3. 3Wonder Full73.75%
  4. 4Claude Opus 5.570.64%
  5. 5Claude Fable 570.43%
  6. 6Wonder70.23%
  7. 7Gemini 4 Argon69.76%
  8. 8MiMo V2.6 Pro68.94%
  9. 9Hy4 Preview68.61%
  10. 10GLM 5.368.54%
  11. 11Muse Spark 1.268.47%
  12. 12Kimi K368.20%
  13. 13Claude Opus 4.868.13%
  14. 14MiMo V2.6 Flash67.59%
  15. 15Claude Sonnet 5.567.19%

Source: Vals AI (Public Benefits 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
Public Benefits Bench 1.1
# ModelScoreCost/taskGatePrudentScore
1Claude Opus 5Anthropic76.93%± 1.10$3.95Passes72
2Claude Fable 5.1Anthropic74.90%± 1.13$7.48Rejectedno score
3Wonder FullMeta73.75%± 1.15$4.64Rejectedno score
4Claude Opus 5.5Anthropic70.64%± 1.19$5.29Passes63
5Claude Fable 5Anthropic70.43%± 1.19$4.59Rejectedno score
6WonderMeta70.23%± 1.19$2.46Rejectedno score
7Gemini 4 ArgonGoogle69.76%± 0.00$2.79Passes79
8MiMo V2.6 ProXiaomi68.94%± 1.20$0.08Rejectedno score
9Hy4 PreviewTencent68.61%± 1.21$0.25Rejectedno score
10GLM 5.3Z.ai68.54%± 1.21$0.95Rejectedno score
11Muse Spark 1.2Meta68.47%± 1.21$0.59Rejectedno score
12Kimi K3Moonshot AI68.20%± 1.21$0.91Rejectedno score
13Claude Opus 4.8Anthropic68.13%± 1.21$1.89Passes62
14MiMo V2.6 FlashXiaomi67.59%± 1.22$0.03Rejectedno score
15Claude Sonnet 5.5Anthropic67.19%± 1.22$3.76Passes62
16Qwen 3.8 MaxAlibaba67.12%± 1.22$0.99Not yet assessedno score
17Grok 4.6xAI (SpaceXAI)66.85%± 1.23$0.92Rejectedno score
18GPT-5.6 SolOpenAI66.51%± 1.23$8.85Passes57
19Ember-1Fireworks AI66.17%± 1.23$1.02Not yet assessedno score
20Claude Sonnet 5Anthropic66.03%± 1.23$1.29Passes59
21Grok 4.7xAI (SpaceXAI)65.63%± 1.24$2.25Rejectedno score
22Gemini 3.8 FlashGoogle65.29%± 1.24$1.01Passes62
23DeepSeek V4.1 FlashDeepSeek64.28%± 1.25$0.07Rejectedno score
24MiniMax-M3MiniMax64.14%± 1.25$0.25Rejectedno score
25Claude Haiku 5.5Anthropic63.19%± 1.25$0.53Passes59
26DeepSeek V4 Pro (2026-04-23)DeepSeek62.92%± 1.26$0.46Rejectedno score
27Claude Sonnet 4.6Anthropic62.45%± 1.26$1.11Passes55
28GPT-5.6 TerraOpenAI62.38%± 1.26$1.20Passes53
29GLM 5.1Z.ai61.84%± 1.26$0.34Rejectedno score
30Muse Spark 1.1Meta61.30%± 1.27$0.28Rejectedno score
31GPT-5.6 LunaOpenAI61.16%± 1.27$0.27Passes56
32GPT 5.5OpenAI60.89%± 1.27$3.97Passes51
33Gemini 3.5 FlashGoogle59.47%± 1.28$1.11Passes47
34Inkling SmallThinking Machines59.41%± 1.28$0.10Rejectedno score
35GPT-6.1 SolOpenAI59.34%± 1.28$0.69Passes59
36InklingThinking Machines58.46%± 1.28$0.37Rejectedno score
37Step 5 PreviewStepFun58.12%± 1.28$0.50Rejectedno score
38GPT-6 LunaOpenAI57.65%± 1.28$0.22Passes54
39=GPT-6 SolOpenAI56.63%± 1.29$2.49Passes45
39=Kimi K2.6Moonshot AI56.63%± 1.29$0.91Rejectedno score
41Gemini 3.6 FlashGoogle56.56%± 1.29$0.39Passes46
42Claude Haiku 4.5 (Nonthinking)Anthropic54.26%± 1.30$0.22Passesno score
43Gemini 3.1 Pro Preview (02/26)Google53.79%± 1.30$0.38Rejectedno score
44Gemini 3.5 Flash LiteGoogle51.83%± 1.30$0.10Passes38
45Grok 4.3xAI (SpaceXAI)51.69%± 1.30$0.26Rejectedno score
46Ling 3.0 FlashAnt Group50.74%± 1.30$0.05Rejectedno score
47Mercury 2.5Inception Labs45.26%± 1.29$0.07Not yet assessedno score
48Grok 4.1 Fast (Reasoning)xAI (SpaceXAI)44.79%± 1.29$0.03Rejectedno score
49Laguna M.1Poolside43.84%± 1.29$0.20Rejectedno score
50Laguna XS.2Poolside41.41%± 1.28$0.11Rejectedno score

Source: Vals AI (Public Benefits 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 1.1 · Measured 12 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.