ResearchLLM Verifications Leaderboard

Data as of · 78 models, 80 questions · reviewed by Suhas Bhairav · methodology

Subject · tested 27 Oct 2026

Best AI models for law

Scores on the 6 law questions in the suite (2 from SuperGPQA, 4 from MMLU-Pro). That is a small sample: the bars show each score's 95% interval, and overlapping bars mean the models can't be told apart.

Questions6of 80 in the suite
Top score6/6Gemini 3.8 Flash, Gemini 4 Argon
Tied with the leader54overlapping 95% intervals
Median score50.0%of 72 models

Law: score by model

6 questions · the bar shows each score's 95% interval

Law leaderboard
#ModelCorrectScore95% interval
1Gemini 3.8 FlashGoogle6 / 6100.0%61–100%
1Gemini 4 ArgonGoogle · statistically tied with #16 / 6100.0%61–100%
3Claude Fable 5.1Anthropic · statistically tied with #15 / 683.3%44–97%
3o3OpenAI · statistically tied with #15 / 683.3%44–97%
3Claude Opus 5.5Anthropic · statistically tied with #15 / 683.3%44–97%
6InklingThinking Machines · statistically tied with #14 / 666.7%30–90%
6Muse WonderMeta · statistically tied with #14 / 666.7%30–90%
6Grok 4.7SpaceXAI · statistically tied with #14 / 666.7%30–90%
6GPT-5.6 TerraOpenAI · statistically tied with #14 / 666.7%30–90%
6GPT-5.3-CodexOpenAI · statistically tied with #14 / 666.7%30–90%
6Hy3Tencent · statistically tied with #14 / 666.7%30–90%
6GLM 5.3 FlashZ.ai · statistically tied with #14 / 666.7%30–90%
6Qwen3 Coder NextQwen · statistically tied with #14 / 666.7%30–90%
6Inkling SmallThinking Machines · statistically tied with #14 / 666.7%30–90%
6Gemini 3.5 Flash LiteGoogle · statistically tied with #14 / 666.7%30–90%
6DeepSeek V4 Pro 0813DeepSeek · statistically tied with #14 / 666.7%30–90%
6Muse Spark OpenMeta · statistically tied with #14 / 666.7%30–90%
6GLM 5.3Z.ai · statistically tied with #14 / 666.7%30–90%
6DeepSeek V4.1 Flash 1017DeepSeek · statistically tied with #14 / 666.7%30–90%
6GPT-5.4OpenAI · statistically tied with #14 / 666.7%30–90%
6GPT-6 SolOpenAI · statistically tied with #14 / 666.7%30–90%
6Qwen3.8 Max (0902)Qwen · statistically tied with #14 / 666.7%30–90%
23Kimi K2.7 CodeMoonshotAI · statistically tied with #13 / 650.0%19–81%
23Command ACohere · statistically tied with #13 / 650.0%19–81%
23Qwen3.7 PlusQwen · statistically tied with #13 / 650.0%19–81%
23GPT-6 AstraOpenAI · statistically tied with #13 / 650.0%19–81%
23GPT-6.1 TerraOpenAI · statistically tied with #13 / 650.0%19–81%
23GPT-6.1 Terra ProOpenAI · statistically tied with #13 / 650.0%19–81%
23Mercury 2.5Inception · statistically tied with #13 / 650.0%19–81%
23MiMo-V2.6-FlashXiaomi · statistically tied with #13 / 650.0%19–81%
23MiniMax M3MiniMax · statistically tied with #13 / 650.0%19–81%
23Qwen3.8 27BQwen · statistically tied with #13 / 650.0%19–81%
23Gemini 3.1 Pro PreviewGoogle · statistically tied with #13 / 650.0%19–81%
23GPT-6.1 SolOpenAI · statistically tied with #13 / 650.0%19–81%
23GPT-6 LunaOpenAI · statistically tied with #13 / 650.0%19–81%
23MiMo-V2.6-ProXiaomi · statistically tied with #13 / 650.0%19–81%
23Llama 4 MaverickMeta · statistically tied with #13 / 650.0%19–81%
23Kimi K3.1 ProMoonshotAI · statistically tied with #13 / 650.0%19–81%
23Claude Sonnet 5.5Anthropic · statistically tied with #13 / 650.0%19–81%
40Step 5 PreviewStepFun · statistically tied with #12 / 633.3%10–70%
40Mistral Medium 3.5Mistral · statistically tied with #12 / 633.3%10–70%
40Qwen3.8 2.4T A95BQwen · statistically tied with #12 / 633.3%10–70%
40Llama 4 ScoutMeta · statistically tied with #12 / 633.3%10–70%
40Hermes 4 405BNous · statistically tied with #12 / 633.3%10–70%
40gpt-oss-20bOpenAI · statistically tied with #12 / 633.3%10–70%
40Phi 4Microsoft · statistically tied with #12 / 633.3%10–70%
40Granite 4.2 8BIBM · statistically tied with #12 / 633.3%10–70%
40Ministral 3 14B 2512Mistral · statistically tied with #12 / 633.3%10–70%
40Mistral Small 4Mistral · statistically tied with #12 / 633.3%10–70%
40DeepSeek V4 Flash Vision ExpDeepSeek · statistically tied with #12 / 633.3%10–70%
40Trinity Large ThinkingArcee AI · statistically tied with #12 / 633.3%10–70%
40LongCat 2.0Meituan · statistically tied with #12 / 633.3%10–70%
40Mistral Large 4Mistral · statistically tied with #12 / 633.3%10–70%
40Claude Haiku 5.5Anthropic · statistically tied with #12 / 633.3%10–70%
40MiniMax M3.1 FlashMiniMax · statistically tied with #12 / 633.3%10–70%
56Gemma 4 26B A4BGoogle1 / 616.7%3–56%
56Gemma 4 31BGoogle1 / 616.7%3–56%
56Ling 3.1 FlashinclusionAI1 / 616.7%3–56%
56Qwen3.5 397B A17BQwen1 / 616.7%3–56%
56Granite 4.0 MicroIBM1 / 616.7%3–56%
56Ministral 3 3B 2512Mistral1 / 616.7%3–56%
56Nova Micro 1.0Amazon1 / 616.7%3–56%
56Ministral 3 8B 2512Mistral1 / 616.7%3–56%
56Nemotron 3.5 LightningNVIDIA1 / 616.7%3–56%
56Llama 3.3 70B InstructMeta1 / 616.7%3–56%
56Solar Pro 4Upstage1 / 616.7%3–56%
56Qwen3.6 35B A3BQwen1 / 616.7%3–56%
56Nemotron 3 UltraNVIDIA1 / 616.7%3–56%
56gpt-oss-120bOpenAI1 / 616.7%3–56%
70Reka Flash 3rekaai0 / 60.0%0–39%
70Nemotron 3.5 Super VLNVIDIA0 / 60.0%0–39%
70Solar Mini 4Upstage0 / 60.0%0–39%

How hard are these questions?

Questions by the share of 72 scored models that got them right

  1. 1No model right0 questions
  2. 2Under 25% of models2 questions
  3. 325–50% of models1 question
  4. 450–75% of models2 questions
  5. 575% or more, not all1 question
  6. 6Every model right0 questions

Science →Engineering →Medicine →Economics →Humanities & social science →

Law: frequently asked questions

Which AI model is best at law?
Gemini 3.8 Flash and Gemini 4 Argon answered the most law questions correctly: 6 of 6. With only 6 questions, 54 other models are statistically tied with the leader, so treat this ranking as a rough guide.
Where do the law questions come from?
2 from SuperGPQA (ODC-BY 1.0) and 4 from MMLU-Pro (MIT). Every question is multiple choice with ten options and one correct answer.
How hard are the law questions?
0 of the 6 questions were answered correctly by none of the 72 scored models, and 0 by all of them.