VLM-SABRE: leaderboard
Vision-language shortcut and bias probes: macro average over four subsets that each strip a different cue a model may be leaning on instead of looking at the image; accuracy %, and scores run low by design.
Metric: Macro-average Accuracy (%). Source: zesearch.github.io. Saturation forecast: Around September 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.6 | 23.3 | #104 |
| 2 | GPT-5.4 | 18 | #79 |
| 3 | Grok 4.3 | 17.8 | #188 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vlm-sabre · How It Works · Data refreshed daily, snapshot 2026-10-20.