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
GPT-6 AstraOpenAI
96.0%
avg
2
Gemini 3.1 ProGoogle DeepMind
95.5%
avg
3
GPT-6.1 SolOpenAI
95.4%
avg
4
Gemini 3.8 FlashGoogle DeepMind
95.3%
avg
5
Meta: Muse Spark 1.3Meta AI
93.8%
avg
6
Claude Fable 5.1Anthropic
93.7%
avg
7
Qwen3.8-2.4T-A95BAlibaba / Qwen
93.5%
avg
8
Kimi K3Moonshot AI
92.9%
avg
9
MiniMax-M3MiniMax
92.9%
avg
10
Qwen3.8-MaxAlibaba / Qwen
92.6%
avg
11
Inkling SmallThinking Machines Lab
92.5%
avg
12
Nanbeige4.3-33BNanbeige
92.3%
avg
13
Tencent Hy4 PreviewTencent
92.3%
avg
14
OrcaCyber Nano 1.0OrcaRouter
92.3%
avg
15
InklingThinking Machines Lab
92.2%
avg
16
Solar Pro 4Upstage
92.2%
avg
17
MAI-Thinking-1Microsoft
91.9%
avg
18
Qwen3.8-Flash-NextAlibaba / Qwen
91.7%
avg
19
GLM-5.3Zhipu AI
91.7%
avg
20
Qwen3.8 FlashAlibaba / Qwen
91.7%
avg
21
GPT-5.3-CodexOpenAI
91.5%
avg
22
DeepSeek V4 Flash Vision ExpDeepSeek
91.3%
avg
23
GLM-5.3-FlashZhipu AI
91.2%
avg
24
Hy3Tencent
90.4%
avg
25
Muse Spark 1.2Meta AI
90.4%
avg
26
Qwen3.5 397B A17BAlibaba / Qwen
90.3%
avg
27
GPT-6 LunaOpenAI
90.0%
avg
28
Qwen3.7 PlusAlibaba / Qwen
90.0%
avg
29
NVIDIA Nemotron 3.5 Super VL 120B-A12B NVFP4NVIDIA
89.9%
avg
30
Kimi K2.7 CodeMoonshot AI
89.6%
avg
31
Qwen3.6 35B A3BAlibaba / Qwen
89.3%
avg
32
Kimi K2 ThinkingMoonshot AI
89.2%
avg
33
Mercury 3Inception
89.2%
avg
34
Qwen3.8-27BAlibaba / Qwen
89.2%
avg
35
Muse GlimmerMeta AI
89.1%
avg
36
Muse Spark OpenMeta AI
89.0%
avg
37
Raon-1 mini previewKrafton
88.8%
avg
38
Seed-2.0-LiteByteDance
88.8%
avg
39
Gemma 4 31BGoogle DeepMind
88.0%
avg
40
K-EXAONE 2.0 (750B-A37B)LG AI Research
87.3%
avg
41
Nemotron 3 UltraNVIDIA
87.0%
avg
42
Gemini 3.1 Flash LiteGoogle DeepMind
86.9%
avg
43
Ling 3.0 Flash VLinclusionAI
86.2%
avg
44
ERNIE 5.0Baidu AI
86.0%
avg
45
Ring-2.6-1TinclusionAI
85.7%
avg
46
Gemma 4 26B A4BGoogle DeepMind
85.3%
avg
47
NVIDIA Nemotron-3-Super-120B-A12BNVIDIA
84.7%
avg
48
GLM 5 TurboZhipu AI
84.7%
avg
49
Seed-2.0-MiniByteDance
84.2%
avg
50
GPT-5.5 InstantOpenAI
83.4%
avg
51
GPT-5.4 miniOpenAI
83.1%
avg
52
MiMo-V2-OmniXiaomi
82.8%
avg
53
Qwen3.5-OmniAlibaba / Qwen
82.6%
avg
54
Qwen3.5-9BAlibaba / Qwen
81.7%
avg
55
GPT-5.4 NanoOpenAI
81.7%
avg
56
Step-3.7-FlashStepFun
80.9%
avg
57
GLM-5V-TurboZhipu AI
80.9%
avg
58
Mistral Medium 3.5Mistral AI
80.5%
avg
59
Mistral Small 4Mistral AI
80.3%
avg
60
ZAYA1-8BZyphra
80.0%
avg
61
Gemma 4 12B UnifiedGoogle DeepMind
78.8%
avg
62
NVIDIA Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4NVIDIA
78.0%
avg
63
LongCat 2.0Meituan
78.0%
avg
64
IBM Granite 4.2 (30B)IBM
76.8%
avg
65
Gemma 4 12BGoogle DeepMind
76.4%
avg
66
Trinity Large ThinkingArcee AI
76.3%
avg
67
Qwen3.5-4BAlibaba / Qwen
76.2%
avg
68
North Mini CodeCohere
75.7%
avg
69
Nemotron 3.5 LightningNVIDIA
75.6%
avg
70
Ling-2.6-1TinclusionAI
75.2%
avg
71
Granite 4.2 8BIBM
74.9%
avg
72
LFM2.5-24B-A2BLiquid AI
74.4%
avg
73
DiffusionGemmaGoogle DeepMind
69.9%
avg
74
Magistral SmallMistral AI
67.7%
avg
75
Mellum2-12B-A2.5B-ThinkingJetBrains
58.0%
avg
76
LFM2.5-2.6BLiquid AI
53.8%
avg
77
Gemma 4 E4B Instruction-TunedGoogle DeepMind
50.5%
avg
78
Nemotron-3-Nano-Omni-30B-A3BNVIDIA
46.9%
avg
79
Granite 4.1 30BIBM
45.8%
avg
80
Qwen3.5-2BAlibaba / Qwen
45.6%
avg
81
Gemma 4 E2BGoogle DeepMind
40.5%
avg
82
Granite 4.1 3BIBM
31.4%
avg
83
LFM2.5-VL-450MLiquid AI
25.7%
avg
84
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 search for the answer still fail.
  • AIME 2026: the American Invitational Math Exam, 2026 edition. 30 hard problems with integer answers. Its problems are recent, so they are less likely to be in training data.
  • AIME 2025: the same format, one year earlier. Used alongside 2026 for a more stable picture of math reasoning capability.

Each score links to its source on the benchmarks leaderboard; scores from independent evaluators (Vals AI, Artificial Analysis) and from developers' own announcements are both used, and developers' figures are often higher.

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