Best LLM models for coding

Coding benchmark results. SWE-bench Verified prefers shared-harness results and falls back to provider reports, identified per score. Terminal-Bench prefers native 4.0, then shared 4.0, then legacy 2.1, with each score labeled. Missing scores mean no result is reported.

Model	DeepSWE ↗v1.1	SWE-bench Verified ↗Shared harness preferred; provider reports otherwise	Terminal-Bench ↗4.0 preferred; 2.1 fallback	LiveBench: Coding ↗2.0 / 2026-10-25
Claude Opus 5Anthropic	74%	—	53.9%4.0 native	—
Gemini 3.8 FlashGoogle	74%	—	19.1%4.0 native	82.2%
GPT-6 AstraOpenAI	74%	—	58.2%4.0 native	93.7%
GPT-5.6 SolOpenAI	73%	—	37.3%4.0 native	—
Claude Fable 5Anthropic	70%	—	44.5%4.0 native	—
GLM-5.3Z.ai	69%	—	41.8%4.0 native	—
Kimi K3Moonshot AI	69%	—	12.6%4.0 shared	59.3%
GPT-5.5OpenAI	67%	—	14.6%4.0 shared	—
Claude Opus 4.8Anthropic	59%	—	23.6%4.0 native	—
Muse Spark 1.2Meta	55%	—	7.1%4.0 shared	—
Claude Sonnet 5Anthropic	54%	—	12.4%4.0 native	—
Claude Fable 5.1Anthropic	—	—	57.9%4.0 native	97.7%
Claude Haiku 5.5Anthropic	—	—	32.8%4.0 shared	91.4%
Claude Opus 4.5Anthropic	—	76.8%Shared harness	—	—
Claude Opus 4.6Anthropic	—	75.6%Shared harness	—	—
Claude Opus 5.5Anthropic	—	—	64.8%4.0 native	96.3%
Claude Sonnet 4.5Anthropic	—	71.4%Shared harness	0%4.0 shared	—
Claude Sonnet 4.6Anthropic	—	—	3%4.0 shared	—
Claude Sonnet 5.5Anthropic	—	—	61.8%4.0 native	94.7%
DeepSeek R1 (Jan 2025)DeepSeek	—	49.2%Provider report	0%4.0 shared	—
DeepSeek V4 FlashDeepSeek	—	—	3%4.0 shared	—
DeepSeek V4 Flash 0420DeepSeek	—	79%Provider report	3%4.0 shared	—
DeepSeek V4 Pro 0424DeepSeek	—	80.6%Provider report	14.6%4.0 shared	—
DeepSeek V4 Pro 0813DeepSeek	—	—	14.1%4.0 shared	—
DeepSeek V4.1 FlashDeepSeek	—	—	26.8%4.0 shared	—
DeepSeek V4.1 Flash 1017DeepSeek	—	—	33.8%4.0 shared	96%
Gemini 3 FlashGoogle	—	75.8%Shared harness	—	—
Gemini 3 ProGoogle	—	69.6%Shared harness	—	—
Gemini 3.1 ProGoogle	—	—	4%4.0 shared	—
Gemini 4 ArgonGoogle	—	—	57.1%4.0 shared	89.7%
Gemini 4 CarbonGoogle	—	—	65.2%4.0 shared	—
GLM-5.1Z.ai	—	—	2%4.0 shared	—
GLM-5.2Z.ai	—	—	1%4.0 shared	—
GLM-5.3-FlashZ.ai	—	—	32.8%4.0 shared	80.6%
GPT-5.2OpenAI	—	72.8%Shared harness	—	—
GPT-5.4 miniOpenAI	—	—	2%4.0 shared	—
GPT-5.6 LunaOpenAI	—	—	11.6%4.0 shared	—
GPT-5.6 TerraOpenAI	—	—	35.4%4.0 shared	86.2%
GPT-6 LunaOpenAI	—	—	12.6%4.0 shared	84.5%
GPT-6.1 SolOpenAI	—	—	58.2%4.0 native	96%
GPT-6.1 TerraOpenAI	—	—	55.6%4.0 shared	94.3%
gpt-oss-120bOpenAI	—	—	0%4.0 shared	—
gpt-oss-20bOpenAI	—	—	0%4.0 shared	—
Grok 4.7xAI	—	—	37.6%4.0 native	84.5%
InklingThinking Machines	—	—	1%4.0 shared	64%
Kimi K2 ThinkingMoonshot AI	—	71.3%Provider report	—	—
Kimi K2.5Moonshot AI	—	70.8%Shared harness	45.7%2.1 legacy	—
Kimi K2.6Moonshot AI	—	80.2%Provider report	0.5%4.0 shared	—
Kimi K3.1 ProMoonshot AI	—	—	67.2%4.0 shared	81.8%
MiMo-V2.6-ProXiaomi	—	—	34.8%4.0 shared	81.5%
MiniMax M2.7MiniMax	—	—	0%4.0 shared	—
MiniMax M3MiniMax	—	80.5%Provider report	2%4.0 shared	—
MiniMax M3.1 FlashMiniMax	—	—	25.8%4.0 shared	83.9%
Muse Spark 1.1Meta	—	—	6.1%4.0 shared	—
Muse Spark 1.3Meta	—	—	14.5%4.0 native	90.2%
Muse Spark OpenMeta	—	—	8.1%4.0 shared	73.6%
Muse Wonder 1.0Meta	—	—	52%4.0 shared	86.2%
Nemotron 3.5 Super VLNVIDIA	—	—	0%4.0 shared	71.3%
Qwen 3.5 9BQwen	—	—	0.5%4.0 shared	—
Qwen3-14BQwen	—	—	0%4.0 shared	—
Qwen3.5-397B-A17BQwen	—	—	0%4.0 shared	—
Qwen3.8 27BQwen	—	—	5.6%4.0 shared	86.8%
Step 5 PreviewStepFun	—	—	33.3%4.0 shared	85.6%

Sorted by DeepSWE (11 reported models), followed by models with no score on that benchmark. SWE-bench Verified prefers shared-harness results and uses labeled provider reports when unavailable. Terminal-Bench prefers native 4.0, then shared 4.0, then legacy 2.1; each score is labeled. Scores across different benchmarks are not directly comparable. Hover a result to see its evaluation settings.