Which model leads MMLU-Redux?

Q: Which model leads MMLU-Redux?

Claude Sonnet 4.5 tops the board at 95.6

MMLU-Redux Leaderboard — 70 Models Ranked

Pro Mode — raw scores, one benchmark under the lensofficial card harness

The field · 69 measured

43.295.6

69 models measured, most on the official card harness. Claude Sonnet 4.5 tops the board at 95.6.

69 measured·Lifecycle Updated Oct 8, 2026Score basis

#◆Vendor$/M in$/M out

Leader — ignites whiteField — fill cools with scoreIndependently verified only0 of 69Bars show 43.2–95.6 · ◆ solid = first-party or better · outlined = arm's-length

01Claude Sonnet 4.5+1 altAnthropic95.6$3.0/M$15.0/M$0.094

02Claude Opus 4.5Anthropic95.63P$5.0/M$25.0/M$0.157

03GPT-5OpenAI95.3$1.3/M$10.0/M$0.059

04Qwen3.7 MaxAlibaba95$2.5/M$7.5/M$0.053

05Qwen3.5 397B A17B+1 altAlibaba94.9$0.60/M$3.6/M$0.022

06Qwen3.7 Plus PreviewAlibaba94.5$0.40/M$1.6/M$0.011

07Qwen3.6 PlusAlibaba94.5$0.50/M$3.0/M$0.019

08Kimi K2 ThinkingMoonshot94.4$0.60/M$2.5/M$0.016

09Claude Opus 4+1 altAnthropic94.23P$15.0/M$75.0/M$0.478

10Qwen3.5 122B A10BAlibaba94$0.40/M$3.2/M$0.019

11Qwen3 VL 235B A22B ReasoningAlibaba93.7$0.40/M$4.0/M$0.023

12Gemma 4 31B+1 altGoogle93.73P$0.17/M$0.40/M$0.003

13DeepSeek-V3.2DeepSeek93.7$0.28/M$0.42/M$0.004

14Claude Sonnet 4+1 altAnthropic93.63P$3.0/M$15.0/M$0.096

15Qwen3.6 27B+1 altAlibaba93.5$0.60/M$3.6/M$0.022

16Qwen3.6 35B A3B+2 altsAlibaba93.3$0.38/M$2.3/M$0.014

17Qwen3.5 35B A3B+1 altAlibaba93.3$0.25/M$2.0/M$0.012

18Qwen3.5 27B+2 altsAlibaba93.2$0.30/M$2.4/M$0.014

19Qwen3 235B A22B Instruct 2507Alibaba93.1$0.23/M$0.92/M$0.006

20DeepSeek-R1+5 altsDeepSeek92.9$2.0/M$4.0/M$0.032

21MiMo V2.5 ProXiaomi92.8$0.43/M$0.87/M$0.007

22Qwen3 Max (Reasoning)Alibaba92.8$1.2/M$6.0/M$0.039

23MiMo V2.5 Pro Base+2 altsXiaomi92.8———

24Gemma 4 26B A4BGoogle92.73P$0.07/M$0.34/M$0.002

25Kimi K2 Instruct+4 altsMoonshot92.7$0.57/M$2.3/M$0.015

26Qwen3 Next 80B A3BAlibaba92.5$0.15/M$1.2/M$0.007

27Qwen3 VL 235B A22B InstructAlibaba92.2$0.40/M$1.6/M$0.011

28Qwen3 VL 32B ReasoningAlibaba91.9$0.16/M$0.64/M$0.004

29DeepSeek-V3.1+2 altsDeepSeek91.8$0.56/M$1.7/M$0.012

30Qwen3.5 9BAlibaba91.1$0.14/M$0.20/M$0.002

31Qwen3 VL 30B A3B ReasoningAlibaba90.9$0.20/M$2.4/M$0.014

32Qwen3 Next 80B A3B InstructAlibaba90.9$0.15/M$1.2/M$0.007

33DeepSeek-V4-Pro-Base+6 altsDeepSeek90.8———

34MiMo V2 Flash Base+2 altsXiaomi90.63P———

35DeepSeek-V3.2-Exp-Base+2 altsDeepSeek90.43P———

36Kimi K2 Base+7 altsMoonshot90.2———

37DeepSeek-V3.1-Base+2 altsDeepSeek903P———

38MiMo V2.5 Base+2 altsXiaomi89.8———

39Qwen3 VL 32B InstructAlibaba89.8$0.16/M$0.64/M$0.004

40DeepSeek-V4-Flash-Base+5 altsDeepSeek89.4———

41Step3.5 Flash Base+1 altStepFun89.23P———

42DeepSeek-V3+9 altsDeepSeek89.1$0.24/M$0.90/M$0.006

43Claude 3.5 Sonnet+4 altsAnthropic88.9$3.0/M$15.0/M$0.101

44Qwen3.5 4BAlibaba88.8$0.03/M$0.15/M$0.001

45Qwen3 VL Thinking (8B)Alibaba88.8$0.18/M$2.1/M$0.013

46Qwen3 14BAlibaba88.6$0.35/M$1.4/M$0.010

47Qwen3 VL 30B A3B InstructAlibaba88.4$0.20/M$0.80/M$0.006

48GPT-4o+4 altsOpenAI88$5.0/M$15.0/M$0.114

49DeepSeek-V3.2-Base+2 altsDeepSeek87.5———

50Qwen3 235B A22B+2 altsAlibaba87.4$0.70/M$2.8/M$0.020

51Hy3-preview-Base+1 altTencent86.86———

52DeepSeek-V3-Base+1 altDeepSeek86.81———

53Qwen2.5 Instruct 72B+1 altAlibaba86.8$0.47/M$0.49/M$0.006

54O1 Mini+3 altsOpenAI86.7———

55GLM-4.5-Base+1 altZ.ai86.56———

56Llama 3.1 Instruct 405BMeta86.2———

57Qwen3 VL 4B (Reasoning)Alibaba86———

58DeepSeek-R1-ZeroDeepSeek85.6———

59Qwen3 VL 8B InstructAlibaba84.9$0.18/M$0.70/M$0.005

60Mistral Large 3Mistral82$0.50/M$1.5/M$0.012

61Qwen3 VL 4B InstructAlibaba81.5———

62DeepSeek-V2.5DeepSeek80.3———

63Qwen3.5 2BAlibaba79.6———

64DeepSeek-V2DeepSeek77.9———

65Qwen 2.5 Coder 32B InstructAlibaba77.5$0.06/M$0.20/M$0.002

66Qwen2.5 Omni 7BAlibaba71$0.10/M$6.8/M$0.048

67Qwen3 4B Base DenseAlibaba61.66———

68Qwen3.5 0.8BAlibaba59.5———

69ERNIE 4.5Baidu43.2———

◆ solid = independently verified · aggregator attested · vendor attributed — outlined = cross-reference · source-attributed. Dimmed ◆ = confidence inferred. “+N alts” = alternate disclosures in the drawer with why-demoted. Prices per million tokens; $/point = input price ÷ score. Every row opens its model dossier.

Verification: 69 models scored · 0 independently verified · 17 vendor cross-reference · 52 vendor-reported · 0 with source disagreement. How these tiers are assigned