SWE Atlas (All Workflows) - Pass^3: leaderboard

SWE Atlas (Scale AI, 2026) evaluates coding agents on 284 tasks in real repositories across three professional workflows, Codebase Q and A, Test Writing and Refactoring, scored by programmatic checks plus must-have rubrics, with three trials per task and each model run in an agent harness. This board is the percent of tasks for which all three trials pass, a consistency measure; higher is better.

Metric: Pass^3 (%): share of tasks where all three trials pass (consistency), over all 284 SWE Atlas tasks (124 Codebase Q&A, 90 Test Writing and 70 Refactoring tasks in real repositories, scored by programmatic checks plus must-have rubrics), three trials per task run through the Harbor framework, each model in its provider's own agent harness (Codex CLI, Claude Code, Gemini CLI) or in mini-SWE-agent; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4xHigh28.9#135
2Claude Opus 4.7xHigh27.1#85
3GPT-5.3 CodexxHigh24.3#124
4Claude Opus 4.6High21.1#113
5Claude Sonnet 4.6High14.4#173
6GLM-511.6#269
7Gemini 3.1 Pro (Preview)High11.3#99
8Kimi K2.59.9#266
9Gemini 3 FlashHigh7.7#175
10MiniMax-M2.56#600

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=swe-atlas-all-workflows-pass-pow-3 · How It Works · Data refreshed daily, snapshot 2026-11-21.