Vals AI Access to Justice (Access to Justice)
We show this table for reference; we do not rank on it.
Can AI help someone navigate a civil legal problem without a lawyer?
Evaluation results published by Vals AI. Published leaderboard aggregates and evaluation settings only. Dataset contents are not included.
Rubric criteria met on Access to Justice — November 10, 2026
We mirror the published rubric criteria met view for Access to Justice. Qwen3.8 Max leads the public snapshot at 76.37%, followed by Claude Sonnet 5.5 (74.94%) and Muse Wonder Full (74.68%). We do not use these results to rank models overall.
Qwen3.8 Max
Alibaba
Claude Sonnet 5.5
Anthropic
Muse Wonder Full
Meta
max reasoning
23 modelsAgenticCurrentDisplay onlyUpdated November 10, 2026
Rubric criteria met table (23 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Qwen3.8 MaxAlibaba | 76.37% | Not reported | Open |
| 2 | Claude Sonnet 5.5Anthropic | 74.94% | Not reported | Closed |
| 3 | Muse Wonder FullMetamax reasoning | 74.68% | Not reported | Closed |
| 4 | Claude Opus 5.5Anthropic | 73.77% | Not reported | Closed |
| 5 | GLM-5.3Z.AImax reasoning | 73.70% | Not reported | Open |
| 6 | Grok 4.7xAIxhigh reasoning | 73.69% | Not reported | Closed |
| 7 | MiniMax M3.1 FlashMiniMax | 73.31% | Not reported | Closed |
| 8 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoning | 73.11% | Not reported | Open |
| 9 | Muse WonderMetamax reasoning | 72.96% | Not reported | Closed |
| 10 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoning | 72.86% | Not reported | Open |
| 11 | Muse Spark 1.3Metaxhigh reasoning | 72.18% | Not reported | Closed |
| 12 | Muse Spark 1.3 MaxMetamax reasoning | 72.15% | Not reported | Unknown |
| 13 | Grok 4.6xAIhigh reasoning | 71.34% | Not reported | Closed |
| 14 | GLM-5.3-FlashZ.AImax reasoning | 71.11% | Not reported | Open |
| 15 | MiMo-V2.6-FlashXiaomi | 70.75% | Not reported | Open |
| 16 | MiMo-V2.6-ProXiaomi | 69.96% | Not reported | Open |
| 17 | MiniMax M3MiniMax | 69.77% | Not reported | Open |
| 18 | Kimi K3.1 ProMoonshot AImax reasoning | 69.47% | Not reported | Open |
| 19 | Gemini 4 ArgonGooglehigh reasoning | 68.48% | Not reported | Closed |
| 20 | Kimi K3Moonshot AImax reasoning | 68.17% | Not reported | Open |
| 21 | Gemini 3.8 FlashGooglehigh reasoning | 67.06% | Not reported | Closed |
| 22 | GPT-6.1 SolOpenAImax reasoning | 65.84% | Not reported | Closed |
| 23 | GPT-5.6 LunaOpenAImax reasoning | 65.56% | Not reported | Closed |
How to read this leaderboard
Operator receipt: 23 sourced rows are currently displayable on this page; the leading published row is Qwen3.8 Max at 76.37%.
Honest limit: Scenarios follow the topic mix legal aid practitioners see, so family law (110) and housing (65) dominate while public benefits has five scenarios. Models were run through their APIs, which approximates but does not reproduce what a chatbot user sees. Answers a provider's content filter blocks count as zero.
How Access to Justice is shown here
We mirror Vals AI's Access to Justice benchmark, published November 10, 2026 with support from NextLadder Ventures. It asks whether a model can act as a reliable first point of contact for someone facing a civil legal problem without a lawyer, the role an intake worker or a court self-help center plays.
Vals simulated 252 conversations from real questions people brought to legal aid organizations, across eight topics led by family law (110 scenarios) and housing (65). Legal aid experts wrote the rubric: a good first answer asks about the person's situation, explains the law accurately and points to the right local help. The score is the share of rubric criteria met. The results stay display only and do not enter model rankings.
Snapshot
The published Access to Justice snapshot places Qwen3.8 Max first at 76.37%. The third row is 1.69 points behind. The broader top-10 range is 3.51 points, so many of the published results sit in a relatively narrow band.
23 models have been evaluated on Access to Justice. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. Access to Justice is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Access to Justice
Year
2026
Tasks
252 simulated legal-aid conversations across 8 topics
Format
Share of expert rubric criteria met
Difficulty
First-contact civil legal help with local referrals
Vals published Access to Justice on November 10, 2026. Each of 252 scenarios starts from a real query brought to a legal aid organization and runs as a simulated conversation; legal aid experts wrote the grading criteria. A good first response asks about the person's situation, gives accurate general legal information, and steers the person to the right local help such as a legal aid program or a court self-help center. The score is the share of rubric criteria met. BenchLM mirrors the published aggregates as display-only external evidence, outside overall and category rankings.
Freshness and provenance
Version
Access to Justice 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Access to Justice measure?
Can AI help someone navigate a civil legal problem without a lawyer?
Which model leads the published Access to Justice snapshot?
Qwen3.8 Max currently leads the published Access to Justice snapshot with 76.37% rubric criteria met. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on Access to Justice?
The November 10, 2026 snapshot contains 23 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.