Coding

SWE-Marathon leaderboard

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap
License: https://creativecommons.org/licenses/by-nc/4.0/

As of 2026-10-13, the highest measured score on SWE-Marathon is 72.7% by Step 5 Preview. 9 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	Step 5 Preview	72.7%	measured
2	Claude Opus 5	60.0%	estimated ± 14.1 pp, low confidence
3	Claude Opus 5.5	59.8%	estimated ± 14.1 pp, low confidence
4	Claude Fable 5.1	59.5%	estimated ± 14.1 pp, low confidence
5	Claude Sonnet 5.5	58.6%	estimated ± 14.1 pp, low confidence
6	Muse Wonder Full	57.5%	measured
7	GLM-5.2	56.2%	estimated ± 14.1 pp, low confidence
8	Kimi K2.7 Code	54.1%	estimated ± 14.1 pp, low confidence
9	Muse Wonder	53.1%	measured
10	MiMo-V2.6-Pro	43.6%	estimated ± 14.1 pp, low confidence
11	MiMo-V2.6-Flash	43.3%	estimated ± 14.1 pp, low confidence
12	GLM-5.3	42.5%	measured
13	Kimi K3	42.0%	measured
14	DeepSeek V4.1 Flash	39.1%	estimated ± 14.1 pp, low confidence
15	Hy4 preview	31.9%	measured

SWE-Marathon leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General