SWE-Marathon v1.2: leaderboard

SWE-Marathon v1.2, the current release: binary task success over up to eight trials per task on 20 scored tasks (15 public sample and 5 held-out), grouped by model, agent and reasoning effort. Kept separate from the archived v1.0 and v1.1 task sets.

Metric: Pass@1 (%). Source: www.swe-marathon.org. Saturation forecast: Rough model projection: around 2026. 13 models tracked. Updated 21 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Fable 5.2Thinking, Max86.79#1
2Claude Opus 5.5Thinking, Max81.25#9
3Claude Sonnet 5.5Thinking, Max80.63#24
4GPT-6.2 AstraThinking, Max77.5#4
5GPT-6 AstraThinking, Max68.75#10
6GPT-6.2 SolThinking, Max66.88#11
7Gemini 4 CarbonThinking, high66.68#7
8Gemini 4 NeonThinking, xhigh62.5#3
9GPT-6.1 SolThinking, Max60.38#15
10Gemini 4 HydrogenThinking, xhigh58.13#16
11Gemini 4 ArgonThinking, High54.46#19
12Claude Haiku 5.5Thinking, Max48.13#75
13Muse WonderThinking, max38.13#35

Interactive version: theaggregate.ai/benchmark?slug=swe-marathon-v1-2 · How It Works · Data refreshed daily, snapshot 2026-11-21.