MCBench: leaderboard

Metric: Average accuracy (%) over the eight category-by-label cells (safe and unsafe scenarios of physical, social, illegal and property harm) of safe-or-unsafe judgments from an image, an audio clip and an utterance, chain-of-thought prompt, GPT-4o judge maps free-form answers to a label, mean of five runs; random 50; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 8 models tracked.

Top models

#ModelScore
1Gemini 2.5 Flash64.4
2Gemma 4 12B (IT)61.9
3Qwen2.5-Omni-7B55.2

Interactive version: theaggregate.ai/benchmark?slug=mcbench · How It Works · Data refreshed daily, snapshot 2026-09-29.