TPSTokens Per Second

Best LLM for Coding in 2026

Models from the last 12 months, ranked by the average of their SWE-bench Verified and LiveCodeBench scores. Scores without a traceable source are left out.

Updated automatically as new models release. Full benchmark leaderboard →

1
GPT-5.6 SolOpenAI
96.2%
avg
2
Claude Fable 5Anthropic
95.0%
avg
3
GPT-5.6 LunaOpenAI
93.0%
avg
4
Ember-1Fireworks Ai
92.2%
avg
5
Kimi K3Moonshot AI
90.3%
avg
6
Claude Opus 4.8Anthropic
88.6%
avg
7
Claude Opus 4.7Anthropic
87.6%
avg
8
Grok 4.5xAI
86.6%
avg
9
Qwen3.7 MaxAlibaba / Qwen
86.0%
avg
10
DeepSeek V4 ProDeepSeek
85.5%
avg
11
Kimi K2.6Moonshot AI
84.9%
avg
12
GPT-5.2OpenAI
84.5%
avg
13
Gemini 3.1 ProGoogle DeepMind
83.7%
avg
14
GLM-5.2Zhipu AI
82.8%
avg
15
GLM-4.7Zhipu AI
81.6%
avg
16
Claude Opus 4.6Anthropic
81.5%
avg
17
GPT-5.4OpenAI
81.2%
avg
18
MiniMax-M2.5MiniMax
80.2%
avg
19
Inkling SmallThinking Machines
80.2%
avg
20
Claude Sonnet 4.6Anthropic
79.8%
avg
21
Claude Sonnet 5Anthropic
79.6%
avg
22
Solar Pro 4Upstage
79.2%
avg
23
Gemini 3 FlashGoogle DeepMind
78.8%
avg
24
Qwen 3.6 PlusAlibaba / Qwen
78.8%
avg
25
Grok 4.20xAI
78.3%
avg
26
Hy3Tencent
78.0%
avg
27
MiMo-V2-ProXiaomi
78.0%
avg
28
GLM 5Zhipu AI
77.8%
avg
29
Qwen3.7 PlusAlibaba / Qwen
77.7%
avg
30
Mistral Medium 3.5Mistral AI
77.6%
avg
31
Kimi K2.5Moonshot AI
76.8%
avg
32
Qwen3.5 397B A17BAlibaba / Qwen
76.4%
avg
33
MiniMax-M2MiniMax
76.0%
avg
34
Muse GlimmerMeta AI
76.0%
avg
35
Claude Opus 4.5Anthropic
75.1%
avg
36
MiniMax-M3MiniMax
75.0%
avg
37
Xing4.0-29B-A4BChina Telecom Ai
75.0%
avg
38
MiMo-V2-OmniXiaomi
74.8%
avg
39
Laguna M.1Poolside
74.6%
avg
40
Muse SparkMeta AI
74.4%
avg
41
Step-3.5-FlashStepFun
74.4%
avg
42
Seed-2.0-LiteByteDance
73.5%
avg
43
MAI-Thinking-1Microsoft
73.5%
avg
44
Qwen3.6 35B A3BAlibaba / Qwen
73.4%
avg
45
GPT-5.4 miniOpenAI
73.0%
avg
46
MAI Code 1.1 FlashMicrosoft
72.6%
avg
47
Qwen3.5-27BAlibaba / Qwen
72.4%
avg
48
Nemotron 3 UltraNVIDIA
71.9%
avg
49
NVIDIA Nemotron 3.5 Super VL 120B-A12B NVFP4NVIDIA
71.1%
avg
50
Laguna XS 2.1Poolside
70.9%
avg
51
Qwen3.6 27BAlibaba / Qwen
70.0%
avg
52
Qwen3.5-35B-A3BAlibaba / Qwen
69.2%
avg
53
Laguna XS.2Poolside
68.2%
avg
54
K-EXAONE 2.0 (750B-A37B)Lg Ai Research
68.2%
avg
55
North Mini Code 1.0Cohere
67.6%
avg
56
Trinity Large ThinkingArcee Ai
63.2%
avg
57
NVIDIA Nemotron-3-Super-120B-A12BNVIDIA
60.5%
avg
58
IBM Granite 4.2 (30B)Ibm
57.0%
avg
59
Nemotron 3.5 LightningNVIDIA
52.8%
avg
60
LFM2.5-24B-A2BLiquid Ai
51.0%
avg
61
Granite 4.2 8BIbm
47.7%
avg

How we rank coding models

We average two benchmarks:

  • SWE-bench Verified: real GitHub issues from popular open-source repos. It tests autonomous software engineering: read the issue, write a fix, pass the test suite.
  • LiveCodeBench: competitive programming problems from LeetCode, Codeforces and AtCoder, collected after model training cutoffs to limit contamination.

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. Rankings update as new scores are added.

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