TPSTokens Per Second

Best LLM for Reasoning in 2026

Ranked by composite reasoning score averaging GPQA, MATH, AIME 2025, and AIME 2024. These benchmarks test graduate-level science, competition mathematics, and multi-step logical reasoning.

Updated automatically as new models release. Full benchmark leaderboard →

1
Grok 4.20xAI
97.0%
avg
2
GPT-5.4OpenAI
95.3%
avg
3
Claude Mythos PreviewAnthropic
94.6%
avg
4
Gemini 3.1 ProGoogle DeepMind
93.7%
avg
5
o3OpenAI
93.5%
avg
6
Kimi K2.6Moonshot AI
93.5%
avg
7
Grok 4xAI
92.3%
avg
8
Gemini 2.5 ProGoogle DeepMind
92.0%
avg
9
o4-miniOpenAI
91.9%
avg
10
Claude Opus 4.6Anthropic
91.4%
avg
11
DeepSeek R1DeepSeek
89.8%
avg
12
Grok 3 minixAI
89.6%
avg
13
Claude Opus 4.5Anthropic
89.3%
avg
14
Qwen3.6 27BAlibaba / Qwen
87.8%
avg
15
Gemma 4 31BGoogle DeepMind
87.6%
avg
16
Qwen3.5-9BAlibaba / Qwen
87.1%
avg
17
Nemotron 3 UltraNVIDIA
87.0%
avg
18
Nemotron-3-Ultra-550B-A55BNVIDIA
87.0%
avg
19
Claude Opus 4Anthropic
86.5%
avg
20
Trinity Large ThinkingArcee Ai
86.3%
avg
21
o1-proOpenAI
86.3%
avg
22
Qwen3.6-35B-A3B-FP8Alibaba / Qwen
86.0%
avg
23
Qwen3.6 35B A3BAlibaba / Qwen
86.0%
avg
24
GPT-5.4 miniOpenAI
85.8%
avg
25
Qwen3.5-27BAlibaba / Qwen
85.5%
avg
26
Gemma 4 26B A4BGoogle DeepMind
85.3%
avg
27
Grok 3xAI
84.9%
avg
28
NVIDIA Nemotron-3-Super-120B-A12BNVIDIA
84.6%
avg
29
Grok 4 minixAI
84.4%
avg
30
Qwen3.5-35B-A3BAlibaba / Qwen
84.2%
avg
31
Claude Sonnet 4.5Anthropic
84.2%
avg
32
ZAYA1-8BZyphra
83.1%
avg
33
Gemma 4 26B A4B ITGoogle DeepMind
82.3%
avg
34
Hy3 PreviewTencent
81.7%
avg
35
GPT-4.1OpenAI
81.5%
avg
36
QwQ-32BAlibaba / Qwen
81.5%
avg
37
Claude Sonnet 4.6Anthropic
80.9%
avg
38
Qwen3 72BAlibaba / Qwen
80.8%
avg
39
Llama 4 MaverickMeta AI
79.3%
avg
40
Gemma 4 12B UnifiedGoogle DeepMind
78.8%
avg
41
MiMo-V2.5-ProXiaomi
76.5%
avg
42
Qwen3.5-4BAlibaba / Qwen
76.2%
avg
43
Sonar Deep ResearchPerplexity AI
73.7%
avg
44
Sonar Reasoning ProPerplexity AI
73.3%
avg
45
Magistral SmallMistral AI
70.7%
avg
46
Claude 3.7 SonnetAnthropic
70.7%
avg
47
Gemma 4 31B IT NVFP4NVIDIA
70.7%
avg
48
Mistral Medium 3Mistral AI
69.0%
avg
49
Sonar ReasoningPerplexity AI
66.6%
avg
50
GPT-4oOpenAI
65.1%
avg
51
DeepSeek V4 ProDeepSeek
64.5%
avg
52
Sonar ProPerplexity AI
62.9%
avg
53
Claude Haiku 4.5Anthropic
61.3%
avg
54
Gemma 4 E4B Instruction-TunedGoogle DeepMind
58.6%
avg
55
Mellum2-12B-A2.5B-ThinkingJetBrains
58.0%
avg
56
DeepSeek V4 FlashDeepSeek
57.4%
avg
57
SonarPerplexity AI
56.0%
avg
58
Step-3.5-Flash-BaseStepFun
54.3%
avg
59
Granite 4.1 30BIbm
45.8%
avg
60
Gemma 4 E2BGoogle DeepMind
44.4%
avg
61
Granite 4.1 8BIbm
42.0%
avg
62
Gemma 4 E2B Instruction-TunedGoogle DeepMind
40.5%
avg
63
LFM2.5-VL-450MLiquid Ai
25.7%
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.
  • MATH — Competition mathematics at AMC/AIME difficulty. Tests multi-step algebraic and geometric reasoning.
  • AIME 2025 — The American Invitational Math Exam, 2025 edition. 30 hard problems, integer answers. Most recent math benchmark — 2025 numbers are resistant to training data contamination.
  • AIME 2024 — Same format, one year earlier. Used alongside 2025 to give 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.