SmellBench - True Positive Component: leaderboard

SmellBench (2026) scores coding-agent CLIs on repairing 65 expert-validated architectural code smells in scikit-learn. This board averages the per-task score over the 11 smells the expert judged genuine, where 1 is a full repair, a severity reduction earns partial credit and a worsened smell or a genuine smell dismissed as a false positive scores below 0; range -1 to 1, higher is better.

Metric: Mean per-task score s (-1 to 1: 1 for a full repair or a correctly flagged false positive, partial credit for a severity reduction, negative for a worsened smell or a genuine smell flagged as false positive) over the 11 smells the expert judged genuine, 65 expert-validated hard-severity PyExamine architectural smells in scikit-learn, off-the-shelf coding agent CLIs with the SmellBench agent skill and GEPA-optimized task packets, default reasoning settings, one run per agent; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.3 Codex0.402#124
2Gemini 3.1 Pro (Preview)0.389#99
3GPT-5.3 Codex Spark0.383
4GPT-5.40.357#135
5GPT-5.1 Codex Max0.304#207
6Claude Haiku 4.50.3#583
7Claude Opus 4.60.285#113
8Claude Sonnet 4.60.278#173
9Mistral Vibe + Devstral 2-0.176
10Gemini 3.1 Flash Lite (Preview)-0.627#417

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

Interactive version: theaggregate.ai/benchmark?slug=smellbench-true-positive-component · How It Works · Data refreshed daily, snapshot 2026-11-21.