SWE-Cycle - Test Generation: leaderboard

SWE-Cycle Test Generation scores code agents on the isolated task to write tests that isolate the issue in a patched codebase, over 489 filtered SWE-bench issues (SWE-Cycle, 2026). The score is the percent of static-review and execution credit an agentic judge awards, higher is better.

Metric: Score (%): the mean of the static and dynamic scores (SWE-Judge (Claude Opus 4.5 agent) static code review and dynamic execution, each scored 0, 1 or 2, normalized to 0-1), on the 489 SWE-Cycle issues (225 SWE-bench Verified, 203 SWE-bench Pro, 61 SWE-bench Multilingual, filtered for contamination, lifecycle complexity and reliable tests); isolated TestGen: write tests that isolate the issue in a patched codebase (90 minutes); agents run in the OpenCode harness in isolated Docker containers; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.688.39#173
2GLM-5.184.36#198
3Kimi K2.581.61#266
4GPT-5.475.41#135
5SWE-Cycle Qwen-3.5 (checkpoint unspecified)74.9
6MiniMax-M2.753.89#539

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

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