TPSTokens Per Second

Best LLM for Reasoning in 2026

Models from the last 12 months, ranked by a composite of GPQA and AIME (2026 and 2025). These benchmarks test graduate-level science, competition mathematics, and multi-step logical reasoning.

Updated automatically as new models release. Full benchmark leaderboard →

1
Gemini 3.1 ProGoogle DeepMind
96.8%
avg
2
Grok 4.20xAI
96.3%
avg
3
GLM-5.2Zhipu AI
95.2%
avg
4
GPT-5.6 SolOpenAI
95.2%
avg
5
GPT-5.2OpenAI
95.2%
avg
6
Claude Mythos PreviewAnthropic
94.6%
avg
7
GPT-5.4OpenAI
94.5%
avg
8
Kimi K3Moonshot AI
93.5%
avg
9
Kimi K2.6Moonshot AI
93.5%
avg
10
Claude Opus 5Anthropic
93.2%
avg
11
Claude Fable 5Anthropic
93.2%
avg
12
GPT-5.5OpenAI
93.2%
avg
13
Grok 4.5xAI
93.1%
avg
14
MiniMax-M3MiniMax
92.9%
avg
15
Qwen3.8-MaxAlibaba / Qwen
92.6%
avg
16
GPT-5.6 TerraOpenAI
92.5%
avg
17
DeepSeek V4 ProDeepSeek
92.3%
avg
18
Gemini 3.5 FlashGoogle DeepMind
92.2%
avg
19
Kimi K2.5Moonshot AI
92.0%
avg
20
Claude Opus 4.8Anthropic
92.0%
avg
21
MAI-Thinking-1Microsoft
91.9%
avg
22
Qwen3.7 MaxAlibaba / Qwen
91.8%
avg
23
GPT-5.3-CodexOpenAI
91.5%
avg
24
Claude Opus 4.7Anthropic
91.4%
avg
25
Claude Opus 4.6Anthropic
91.2%
avg
26
GPT-5.6 LunaOpenAI
91.1%
avg
27
Claude Sonnet 5Anthropic
91.1%
avg
28
GLM-5.1Zhipu AI
91.0%
avg
29
Qwen3.6 27BAlibaba / Qwen
90.9%
avg
30
GLM-4.7Zhipu AI
90.8%
avg
31
Gemini 3 FlashGoogle DeepMind
90.5%
avg
32
Hy3Tencent
90.4%
avg
33
Muse Spark 1.2Meta AI
90.4%
avg
34
Qwen3.5 397B A17BAlibaba / Qwen
90.3%
avg
35
Grok 4.3xAI
90.1%
avg
36
Qwen3.7 PlusAlibaba / Qwen
90.0%
avg
37
Step-3.5-FlashStepFun
89.9%
avg
38
Muse Spark 1.1Meta AI
89.8%
avg
39
Kimi K2.7 CodeMoonshot AI
89.6%
avg
40
Inkling SmallThinking Machines
89.5%
avg
41
DeepSeek V4 FlashDeepSeek
89.4%
avg
42
Qwen3.6 35B A3BAlibaba / Qwen
89.3%
avg
43
Kimi K2 ThinkingMoonshot AI
89.2%
avg
44
Nex-N2-ProNex Agi
89.2%
avg
45
Seed-2.0-LiteByteDance
88.8%
avg
46
Muse SparkMeta AI
88.4%
avg
47
Claude Opus 4.5Anthropic
88.0%
avg
48
Gemma 4 31BGoogle DeepMind
87.5%
avg
49
MiniMax M2.7MiniMax
87.4%
avg
50
K-EXAONE 2.0 (750B-A37B)Lg Ai Research
87.3%
avg
51
Hy3 PreviewTencent
87.2%
avg
52
InklingThinking Machines
87.2%
avg
53
Qwen3.5-9BAlibaba / Qwen
87.1%
avg
54
Nemotron-3-Ultra-550B-A55BNVIDIA
87.0%
avg
55
Nemotron 3 UltraNVIDIA
87.0%
avg
56
MiMo-V2-ProXiaomi
87.0%
avg
57
Gemma 4 31B InstructGoogle DeepMind
86.8%
avg
58
Gemma 4 26B A4BGoogle DeepMind
86.3%
avg
59
ERNIE 5.0Baidu AI
86.0%
avg
60
Qwen3.6-35B-A3B-FP8Alibaba / Qwen
86.0%
avg
61
Ring-2.6-1TInclusionai
85.7%
avg
62
Ling-3.0-flashInclusionai
85.5%
avg
63
Qwen3.5-27BAlibaba / Qwen
85.5%
avg
64
KAT-Coder-Pro V2Kwaipilot
85.5%
avg
65
Claude Sonnet 4.5Anthropic
85.2%
avg
66
MiniMax-M2.5MiniMax
84.8%
avg
67
GLM 5 TurboZhipu AI
84.7%
avg
68
NVIDIA Nemotron-3-Super-120B-A12BNVIDIA
84.6%
avg
69
Qwen3.5-35B-A3BAlibaba / Qwen
84.2%
avg
70
Mercury 2Inception
84.0%
avg
71
GPT-5.5 InstantOpenAI
83.4%
avg
72
MiMo-V2-OmniXiaomi
82.8%
avg
73
Qwen3.5-OmniAlibaba / Qwen
82.6%
avg
74
Gemini 3.1 Flash Lite PreviewGoogle DeepMind
82.2%
avg
75
GLM 5Zhipu AI
82.0%
avg
76
GPT-5.4 NanoOpenAI
81.7%
avg
77
Step-3.7-FlashStepFun
80.9%
avg
78
GLM-5V-TurboZhipu AI
80.9%
avg
79
Mistral Medium 3.5Mistral AI
80.5%
avg
80
Mistral Small 4Mistral AI
80.3%
avg
81
ZAYA1-8BZyphra
80.0%
avg
82
Grok 4 minixAI
80.0%
avg
83
Qwen3-MaxAlibaba / Qwen
79.0%
avg
84
Gemma 4 12B UnifiedGoogle DeepMind
78.8%
avg
85
GLM-4.6Zhipu AI
78.6%
avg
86
GPT-5.4 miniOpenAI
78.5%
avg
87
LongCat 2.0Meituan
78.0%
avg
88
NVIDIA Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4NVIDIA
78.0%
avg
89
MiniMax-M2MiniMax
77.8%
avg
90
Claude Sonnet 4.6Anthropic
76.8%
avg
91
Gemma 4 12BGoogle DeepMind
76.4%
avg
92
Trinity Large ThinkingArcee Ai
76.3%
avg
93
Qwen3.5-4BAlibaba / Qwen
76.2%
avg
94
North Mini CodeCohere
75.7%
avg
95
Ling-2.6-1TInclusionai
75.2%
avg
96
DiffusionGemma 26B A4B ITGoogle DeepMind
71.2%
avg
97
Gemma 4 31B IT NVFP4NVIDIA
70.7%
avg
98
Magistral SmallMistral AI
67.7%
avg
99
Kimi K2 0905Moonshot AI
67.0%
avg
100
MiMo-V2.5-ProXiaomi
66.7%
avg
101
Claude Haiku 4.5Anthropic
65.4%
avg
102
Ling-2.6-flashInclusionai
59.3%
avg
103
Mellum2-12B-A2.5B-ThinkingJetBrains
58.0%
avg
104
Mistral Medium 3Mistral AI
54.4%
avg
105
Mistral Large 3Mistral AI
53.0%
avg
106
Gemma 4 E4B Instruction-TunedGoogle DeepMind
50.5%
avg
107
Devstral 2Mistral AI
48.0%
avg
108
Nemotron-3-Nano-Omni-30B-A3BNVIDIA
46.9%
avg
109
Granite 4.1 30BIbm
45.8%
avg
110
Qwen3.5-2BAlibaba / Qwen
45.6%
avg
111
Granite 4.1 8BIbm
42.0%
avg
112
Step-3.5-Flash-BaseStepFun
41.7%
avg
113
Gemma 4 E2B Instruction-TunedGoogle DeepMind
39.5%
avg
114
Granite 4.1 3BIbm
31.4%
avg
115
Mistral SabaMistral AI
27.7%
avg
116
LFM2.5-VL-450MLiquid Ai
25.7%
avg
117
LFM2.5-230MLiquid Ai
25.4%
avg

What makes a good reasoning model?

Reasoning benchmarks test whether a model can solve multi-step problems requiring planning, logic, and domain knowledge — not just pattern matching or retrieval.

  • GPQA (Diamond) — Questions written by PhD-level experts in biology, chemistry, and physics. Designed so that non-experts who Google the answer still fail. The gold standard for deep scientific reasoning.
  • AIME 2026 — The American Invitational Math Exam, 2026 edition. 30 hard problems, integer answers. The current math benchmark labs report — fresh numbers are resistant to training-data contamination.
  • AIME 2025 — Same format, one year earlier. Used alongside 2026 for a more stable picture of math reasoning capability.

For tasks involving complex analysis, research, legal and financial reasoning, or scientific work, a high GPQA score is the best predictor of real-world performance.

Also see: Best Coding LLM, Best Cheap LLM, Compare any two models.