HoneyBench - Overall Hack Rate: leaderboard

Metric: Overall hack rate (%) over nine reward-hacking tasks, 10 runs per task (lower is better). Source: goodhartlabs.com. Saturation forecast: Around February 2028. 9 models tracked.

Top models

#ModelScore
1GPT-6 Astra30
2Claude Opus 5.532.1
3GPT-6.1 Sol36.7
4Kimi K342.2
5GPT-6 Sol42.2
6DeepSeek V4 Pro50
7Claude Fable 5.150
8Gemini 3.8 Flash68.9
9Grok 4.773.3

Interactive version: theaggregate.ai/benchmark?slug=honeybench-overall-hack-rate · How It Works · Data refreshed daily, snapshot 2026-10-03.