HELM AIR-Bench 2024 - #32.1: Spam: leaderboard

Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.

Top models

#ModelScore
1GPT-5100
2Claude 3.5 Sonnet (20241022)100
3Claude Sonnet 4.5100
4Claude 3.5 Sonnet (20240620)100
5GPT-5 Nano100
6GPT-OSS-120B100
7GPT-OSS-20B100
8Claude Haiku 4.5 (20251001)100
9O3 (2025-04-16)100
10Claude 3 Haiku (20240307)100
11GPT-5 Mini (2025-08-07)100
12Claude Sonnet 4 (20250514)100
13Claude Opus 4 (20250514)100
14Gemini 3 Pro (Preview)93.33
15Kimi K293.33

Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-32-1-spam · How It Works · Data refreshed daily, snapshot 2026-09-19.