MusICA-MetaBench (ChoraleBricks): leaderboard

Metric: Accuracy (%; 300 five-option music-perception questions generated from ChoraleBricks MusicXML, 100 each from audio recordings, sheet-music images and ABC notation, one benchmark instance; none-of-the-above is correct 20% of the time; random choice 20%). Source: arxiv.org. Saturation forecast: Around September 2027. 5 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)59
2Gemini 3.1 Flash Lite (Preview)45.7
3Gemini 2.5 Flash Lite41
4Gemini 2.0 Flash Lite40.3

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