SO-Arena v1.0 - Reliability: leaderboard

SO-Arena v1.0 native reliability. Stability measures consistency under option reordering, null discrimination measures missing-answer handling, and reliability combines the two. The published JSON supplies these exact values; StartLux null scores are derived from rounded reliability values. Native model inference paths and API interfaces remain separate. Metric: Native reliability (%). Higher is better.

Metric: Native reliability (%). Source: raw.githubusercontent.com. Saturation forecast: Rough model projection: around 2026. 19 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1JevAny-27B84.11
2Jev 1.14.083.45
3Sieve-31B-v381.31
4KnowLine-31B-Gen181.01
5Kev 27B80.78
6StartLux-v1.1-4B-AutoThink80.5
7StartLux-v1.1-4B-Instant78.9
8GLiDE v1.277.57
9PPLX-Decider-v1.275.29
10Clef 27B73.21
11Clef flash 9.5B71
12JevAny-4B70.54
13PPLX-Decider-Flash-v1.267.92
14GPT-6 LunaSO-Arena run: Decisions API; model=gpt-6-luna61.05#99
15H2O-Lightning-4B60.91

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