Skip to content

Benchmarks · Vals AI

GPQA Diamond (Vals-run)

General reasoning benchmark: graduate-level science multiple-choice questions (run by Vals; "Archived Benchmark", saturated, no new models; last update 1 September 2026). Not legal.

Models measured
109
Top
95.45 %Gemini 3.1 Pro Preview (02/26)within the standard error of ±1.05
Version
1
Measured
1 September 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. 1Gemini 3.1 Pro Preview (02/26)95.45%
  2. 2GPT-5.6 Sol95.20%
  3. 3Grok 4.694.70%
  4. 4Gemini 3.8 Flash94.44%
  5. 5Gemini 3.7 Flash93.94%
  6. 6Qwen 3.8 Max93.69%
  7. 7=Claude Fable 5.193.43%
  8. 7=Claude Opus 593.43%
  9. 7=Gemini 3.6 Flash93.43%
  10. 10=Claude Fable 593.18%
  11. 10=GPT 5.593.18%
  12. 12=Grok 4.592.93%
  13. 12=Kimi K392.93%
  14. 14=Gemini 3.5 Flash92.68%
  15. 14=MiniMax-M392.68%

Source: Vals AI (GPQA Diamond (Vals-run)) · 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
GPQA Diamond (Vals-run) 1
# ModelScoreCost/taskGatePrudentScore
1Gemini 3.1 Pro Preview (02/26)Google95.45%± 1.05-Rejectedno score
2GPT-5.6 SolOpenAI95.20%± 1.07-Passes57
3Grok 4.6xAI (SpaceXAI)94.70%± 1.13-Rejectedno score
4Gemini 3.8 FlashGoogle94.44%± 1.48-Passes62
5Gemini 3.7 FlashGoogle93.94%± 1.49-Passes58
6Qwen 3.8 MaxAlibaba93.69%± 1.22-Not yet assessedno score
7=Claude Fable 5.1Anthropic93.43%± 1.86-Rejectedno score
7=Claude Opus 5Anthropic93.43%± 1.24-Passes72
7=Gemini 3.6 FlashGoogle93.43%± 1.33-Passes46
10=Claude Fable 5Anthropic93.18%± 1.94-Rejectedno score
10=GPT 5.5OpenAI93.18%± 1.29-Passes51
12=Grok 4.5xAI (SpaceXAI)92.93%± 1.29-Rejectedno score
12=Kimi K3Moonshot AI92.93%± 1.31-Rejectedno score
14=Gemini 3.5 FlashGoogle92.68%± 1.46-Passes47
14=MiniMax-M3MiniMax92.68%± 1.44-Rejectedno score
16=Claude Opus 4.8Anthropic92.42%± 1.98-Passes62
16=DeepSeek V4 Pro (0813)DeepSeek92.42%± 2.02-Rejectedno score
18=Gemini 3 Pro (11/25)Google91.67%± 1.39-Rejectedno score
18=GPT 5.2OpenAI91.67%± 1.84-Passesno score
18=GPT 5.4 (xhigh)OpenAI91.67%± 1.91-Passes0
18=GPT-5.6 LunaOpenAI91.67%± 1.74-Passes56
22Grok 4.3xAI (SpaceXAI)91.41%± 1.41-Rejectedno score
23Muse Spark 1.1Meta91.16%± 2.10-Rejectedno score
24GPT-5.6 TerraOpenAI90.91%± 1.91-Passes53
25=Claude Opus 4.7Anthropic90.15%± 1.54-Passes54
25=Qwen 3.7 MaxAlibaba90.15%± 1.50-Not yet assessedno score
27DeepSeek V4 Flash 0731DeepSeek89.90%± 1.69-Rejectedno score
28=Claude Opus 4.6 (Thinking)Anthropic89.65%± 1.19-Passesno score
28=Muse SparkMeta89.65%± 1.55-Rejectedno score
30DeepSeek V4 Pro (2026-04-23)DeepSeek89.39%± 1.65-Rejectedno score
31Kimi K2.6Moonshot AI89.14%± 2.00-Rejectedno score
32=Claude Sonnet 5Anthropic88.89%± 2.22-Passes59
32=Qwen 3.8 27BAlibaba88.89%± 1.58-Rejectedno score
34Grok 4.20 (Reasoning)xAI (SpaceXAI)88.64%± 1.59-Rejectedno score
35=GLM 5.3Z.ai88.13%± 1.85-Rejectedno score
35=Grok 4xAI (SpaceXAI)88.13%± 1.63-Rejectedno score
37Gemini 3 Flash (12/25)Google87.88%± 1.64-Rejectedno score
38=Qwen 3.5 PlusAlibaba87.37%± 1.67-Not yet assessedno score
38=Qwen 3.6 PlusAlibaba87.37%± 1.67-Not yet assessedno score
40InklingThinking Machines87.12%± 1.78-Rejectedno score
41=GPT 5.1OpenAI86.62%± 1.71-Passesno score
41=MiniMax-M2.7MiniMax86.62%± 1.95-Rejectedno score
43GLM 5.3 FlashZ.ai86.36%± 2.13-Rejectedno score
44Nemotron 3 UltraNVIDIA86.11%± 1.74-Rejectedno score
45Claude Opus 4.5 (Thinking)Anthropic85.86%± 2.48-Passesno score
46=Claude Sonnet 4.6Anthropic85.61%± 2.41-Passes55
46=GLM 5.2Z.ai85.61%± 1.77-Rejectedno score
46=GPT 5OpenAI85.61%± 1.76-Passesno score
49Grok 4 Fast (Reasoning)xAI (SpaceXAI)85.35%± 1.79-Rejectedno score
50=Ling 3.0 FlashAnt Group84.85%± 2.20-Rejectedno score
50=Qwen 3 Max ThinkingAlibaba84.85%± 1.80-Not yet assessedno score
52GLM 5.1Z.ai84.52%± 1.82-Rejectedno score
53Grok 4.1 Fast (Reasoning)xAI (SpaceXAI)84.34%± 1.85-Rejectedno score
54=Kimi K2.5Moonshot AI84.09%± 2.13-Passes37
54=o3OpenAI84.09%± 1.86-Passesno score
56Gemini 3.5 Flash LiteGoogle83.84%± 1.85-Passes38
57Inkling SmallThinking Machines83.59%± 2.04-Rejectedno score
58GLM 5Z.ai83.33%± 1.87-Passesno score
59GPT 5.4 MiniOpenAI83.08%± 2.46-Passes25
60Qwen 3.5 FlashAlibaba82.83%± 1.90-Not yet assessedno score
61MiMo V2.5 ProXiaomi82.58%± 2.47-Rejectedno score
62MiniMax-M2.5MiniMax82.07%± 2.26-Passesno score
63Claude Sonnet 4.5 (Thinking)Anthropic81.63%± 2.25-Passesno score
64=Gemini 2.5 Flash Preview (9/25) (Nonthinking)Google81.57%± 2.00-Rejectedno score
64=MiMo V2.5Xiaomi81.57%± 2.05-Rejectedno score
66Gemini 3.1 Flash Lite PreviewGoogle81.06%± 1.97-Not yet assessedno score
67=DeepSeek V3.2 (Thinking)DeepSeek80.30%± 2.00-Passesno score
67=GPT 5 MiniOpenAI80.30%± 2.00-Passesno score
69GLM 4.7Z.ai80.05%± 2.01-Passesno score
70=Claude Opus 4.5 (Nonthinking)Anthropic79.55%± 2.46-Passesno score
70=Qwen 3 MaxAlibaba79.55%± 2.03-Not yet assessedno score
72=GPT OSS 120BOpenAI78.54%± 2.25-Passesno score
72=Kimi K2 ThinkingMoonshot AI78.54%± 2.18-Rejectedno score
72=MiniMax-M2.1MiniMax78.54%± 2.13-Passesno score
75Qwen 3 Max PreviewAlibaba77.78%± 2.09-Not yet assessedno score
76GPT 5.4 NanoOpenAI77.53%± 2.42-Passes29
77Gemini 2.5 Flash Preview (9/25) (Thinking)Google76.52%± 2.15-Rejectedno score
78=Claude Opus 4.1 (Thinking)Anthropic76.26%± 2.16-Rejectedno score
78=DeepSeek V3.2 (Nonthinking)DeepSeek76.26%± 2.16-Passesno score
80Claude Sonnet 4 (Thinking)Anthropic75.00%± 2.19-Rejectedno score
81=GLM 4.6Z.ai74.50%± 2.20-Rejectedno score
81=o4 MiniOpenAI74.50%± 2.19-Passesno score
83=Claude Haiku 4.5 (Thinking)Anthropic72.22%± 2.47-Passes32
83=GLM 4.5Z.ai72.22%± 2.26-Rejectedno score
85Claude Opus 4 (Nonthinking)Anthropic71.72%± 2.26-Rejectedno score
86Kimi K2 InstructMoonshot AI71.46%± 2.27-Rejectedno score
87Gemini 2.5 Flash Lite (9/25) (Thinking)Google70.20%± 2.30-Rejectedno score
88Claude Opus 4.1 (Nonthinking)Anthropic69.95%± 2.30-Rejectedno score
89=Claude Sonnet 4 (Nonthinking)Anthropic69.44%± 2.31-Rejectedno score
89=Llama 4 MaverickMeta69.44%± 2.35-Rejectedno score
91GPT OSS 20BOpenAI68.94%± 2.46-Passesno score
92Mistral Large 3Mistral AI68.43%± 2.37-Passesno score
93GPT 4.1 MiniOpenAI67.93%± 2.35-Passesno score
94GPT 4.1OpenAI65.40%± 2.40-Passesno score
95Grok 4.1 Fast Non-ReasoningxAI (SpaceXAI)65.15%± 2.40-Rejectedno score
96Gemini 2.5 Flash Lite (9/25) (Nonthinking)Google64.14%± 2.41-Rejectedno score
97GPT 5 NanoOpenAI63.38%± 2.46-Passesno score
98Magistral Medium 1.2 (09/2025)Mistral AI62.37%± 2.44-Not yet assessedno score
99Grok 4 Fast (Non-Reasoning)xAI (SpaceXAI)62.12%± 2.43-Rejectedno score
100MiMo V2 FlashXiaomi59.34%± 2.50-Rejectedno score
101Magistral Small 1.2 (09/2025)Mistral AI58.33%± 2.49-Passesno score
102Laguna XS.2Poolside55.05%± 2.50-Rejectedno score
103GPT 4o (2024-11-20)OpenAI53.79%± 2.51-Passesno score
104GPT 4.1 NanoOpenAI50.76%± 2.46-Passesno score
105Command ACohere48.48%± 2.28-Rejectedno score
106Llama 4 ScoutMeta46.97%± 2.49-Rejectedno score
107Mistral Small 3.1 (03/2025)Mistral AI44.19%± 2.47-Passesno score
108Mistral Medium 3.5Mistral AI34.85%± 2.39-Passes23
109Laguna M.1Poolside27.02%± 1.92-Rejectedno score

Source: Vals AI (GPQA Diamond (Vals-run)) · 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 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

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.