Coding

Writing and editing code against a specification, measured without an agent loop. Distinct from agentic coding: these benchmarks score the model's output directly rather than a model-plus-harness system.

Ordered by how many of these benchmarks put a model in their top 10, then by its best placing. Each column keeps its own ranking; nothing is averaged.

Model	In top 10	Artificial Analysis LLM Leaderboard SciCode	Arena — Code Overall
Claude Opus 5.5 anthropic	2/2	#1 66.9%	#1 1815
Claude Fable 5.1 anthropic	2/2	#2 63.1%	#5 1749
Gemini 4 Argon google	2/2	#3 61.8%	#8 1680
Claude Sonnet 5.5 anthropic	2/2	#5 61.0%	#3 1786
Kimi K3 kimi	2/2	#9 59.5%	#10 1658
GPT 6 Astra openai	1/2	#19 56.5%	#2 1788
Claude Fable 5 anthropic	1/2	#4 61.0%	#15 1625
GPT 6.1 Sol openai	1/2	#23 55.8%	#4 1758
MiMo V2.6 Pro xiaomi	1/2	#6 60.9%	#20 1618
Claude Opus 5 anthropic	1/2	#21 56.4%	#6 1695
Gemini 3.7 Flash google	1/2	#7 59.8%	#23 1592
GPT 6 Sol openai	1/2	#16 57.6%	#7 1689
Muse Spark 1.3 meta	1/2	#8 59.7%	#11 1657
Qwen 3.8 Max alibaba	1/2	#29 54.1%	#9 1671
GLM 5.3 zai	1/2	#10 59.0%	#16 1623

A dash means the board does not list that model — not a score of zero. Entries not yet matched to a model are left out here; they still appear on their own board.