SO-Arena v1.0 - Safety: leaderboard

SO-Arena v1.0 Safety domain score on the source-selected1232-task panel. Domain scores average native task credits for static decisions, tool outcomes or web progress, then scale to0–100. The full release has2919tasks; this leaderboard emphasizes the highest difficulty tiers and keeps all untiered sentiment/knowledge tasks. Metric: Native panel task credit mean (%). Higher is better.

Metric: Native panel task credit mean (%). Source: huggingface.co. Saturation forecast: Around February 2027. 20 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Sieve-31B-v378.48
2KnowLine-31B-Gen177.22
3GPT-6 SolMedium75.32#26
4PPLX-Decider-v1.274.68
5GLiDE v1.274.68
6GPT-6 LunaMedium70.25#99
7Jev 1.14.068.99
8JevAny-27B65.19
9Kev 27B63.92
10H2O-Lightning-4B63.92
11StartLux-v1.1-4B-Instant63.29
12StartLux-v1.1-4B-AutoThink62.03
13PPLX-Decider-Flash-v1.260.76
14Clef 27B59.49
15Jeva-4B50.63

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=so-arena-v1-0-safety · How It Works · Data refreshed daily, snapshot 2026-11-16.