GlazyBench - Color Family (5-Shot): leaderboard

GlazyBench (2026) asks whether a model can predict how a ceramic glaze looks after firing from its recipe, given as oxide weight percentages, the unity molecular formula and firing conditions. This board is accuracy in percent at predicting the colour family, one of nine hue-based classes, with five labelled training recipes in the prompt; higher is better.

Metric: Accuracy (%) predicting the fired glaze colour family (9 classes; 4,903 test glazes) from its recipe (oxide weight percentages, UMF formula, cone and atmosphere as text), five-shot with one stratified training example per class in rotation, temperature 0, one label from the allowed set; outputs matching no label are left out of the metric; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GlazyBench DeepSeek-v3 (checkpoint unspecified)17.5
2GPT-4o Mini16.9#1667
3Claude Sonnet 4.515.5#261

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=glazybench-color-family-5-shot · How It Works · Data refreshed daily, snapshot 2026-11-21.