Best LLM for Coding in 2026

Ranked by composite coding score averaging HumanEval, SWE-bench Verified, and LiveCodeBench. All scores from official model cards and technical reports.

Updated automatically as new models release. Full benchmark leaderboard →

Mistral Small 3.1 Mistral AI

88.4%

avg

HumanEval: 88.4%

GPT-4.5 OpenAI

86.5%

avg

HumanEval: 86.5%

Mistral Medium 3 Mistral AI

86.4%

avg

HumanEval: 86.4%

Claude Opus 4.5 Anthropic

85.5%

avg

HumanEval: 90.0%

SWE-bench Verified: 80.9%

Gemini 2.0 Flash Google DeepMind

83.8%

avg

HumanEval: 83.8%

Grok 4 xAI

82.8%

avg

HumanEval: 88.0%

SWE-bench Verified: 81.0%

LiveCodeBench: 79.3%

Phi-4 Microsoft

82.6%

avg

HumanEval: 82.6%

o3 OpenAI

82.1%

avg

HumanEval: 91.6%

SWE-bench Verified: 71.7%

LiveCodeBench: 83.1%

Claude Opus 4.6 Anthropic

80.8%

avg

SWE-bench Verified: 80.8%

o4-mini OpenAI

80.1%

avg

HumanEval: 90.5%

SWE-bench Verified: 68.1%

LiveCodeBench: 81.8%

Claude Sonnet 4.6 Anthropic

79.6%

avg

SWE-bench Verified: 79.6%

Gemini 3.1 Pro Google DeepMind

79.2%

avg

HumanEval: 84.1%

SWE-bench Verified: 80.6%

LiveCodeBench: 73.0%

Claude Haiku 4.5 Anthropic

79.1%

avg

HumanEval: 92.3%

SWE-bench Verified: 73.3%

LiveCodeBench: 71.7%

Claude Opus 4 Anthropic

78.7%

avg

HumanEval: 92.4%

SWE-bench Verified: 72.5%

LiveCodeBench: 71.2%

Gemini 2.5 Pro Google DeepMind

78.4%

avg

HumanEval: 94.3%

SWE-bench Verified: 63.8%

LiveCodeBench: 77.0%

Grok 3 mini xAI

78.3%

avg

HumanEval: 78.3%

QwQ-32B Alibaba / Qwen

78.0%

avg

HumanEval: 91.9%

LiveCodeBench: 64.0%

Claude Sonnet 4.5 Anthropic

77.2%

avg

SWE-bench Verified: 77.2%

Qwen3 72B Alibaba / Qwen

76.3%

avg

HumanEval: 90.0%

LiveCodeBench: 62.6%

Claude 3.7 Sonnet Anthropic

75.8%

avg

HumanEval: 91.0%

SWE-bench Verified: 70.3%

LiveCodeBench: 66.1%

Grok 3 xAI

73.5%

avg

HumanEval: 88.0%

LiveCodeBench: 58.9%

o1-pro OpenAI

70.7%

avg

HumanEval: 92.4%

SWE-bench Verified: 48.9%

DeepSeek-R1 DeepSeek

70.5%

avg

HumanEval: 96.3%

SWE-bench Verified: 49.2%

LiveCodeBench: 65.9%

Gemini 2.5 Flash Google DeepMind

70.5%

avg

HumanEval: 89.2%

SWE-bench Verified: 53.5%

LiveCodeBench: 68.7%

GPT-4.1 OpenAI

68.6%

avg

HumanEval: 92.0%

SWE-bench Verified: 54.6%

LiveCodeBench: 59.3%

DeepSeek R1 (0528)DeepSeek

65.5%

avg

SWE-bench Verified: 57.6%

LiveCodeBench: 73.3%

GPT-4o OpenAI

61.7%

avg

HumanEval: 90.2%

SWE-bench Verified: 33.2%

Llama 4 Maverick Meta AI

54.3%

avg

HumanEval: 87.5%

SWE-bench Verified: 32.0%

LiveCodeBench: 43.4%

How we rank coding models

We average three industry-standard benchmarks:

HumanEval — Function-completion coding tasks in Python. Tests whether a model can write correct, working code from a docstring description. Pass@1 accuracy.
SWE-bench Verified — Real GitHub issues from popular open-source repos. Tests autonomous software engineering: read the issue, write a fix, pass the test suite. The most practical real-world coding benchmark available.
LiveCodeBench — Competitive programming problems from LeetCode, Codeforces, and AtCoder, collected after model training cutoffs to prevent contamination. Harder than HumanEval.

All scores are from official model cards, technical reports, or the HuggingFace Open LLM Leaderboard. Rankings update automatically as new models are released.

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