SADU - Counting: leaderboard

SADU (2026) tests vision-language models on counting and retrieval questions about behavioral, structural and entity-relationship software architecture diagrams. This board is exact-match accuracy in percent on the counting questions across all diagram types; higher is better.

Metric: Exact-match accuracy (%) on the counting questions of SADU across all diagram types (154 software architecture diagrams (behavioral, structural and ER) with 2,431 counting and retrieval questions, temperature 0); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash (Preview)76#152
2Gemini 2.5 Flash72#463
3Gemini 3.1 Flash Lite (Preview)61#417
4Claude Sonnet 4.551#261
5GPT-5 Nano50#906
6Claude Haiku 4.544#583
7Qwen 2.5 VL 32B38#941
8Gemini 2.5 Flash Lite32#964
9Qwen 2.5 VL 7B31#1862
10GPT-4o Mini16#1667
11Qwen 2.5 VL 3B8#2639

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

Interactive version: theaggregate.ai/benchmark?slug=sadu-counting · How It Works · Data refreshed daily, snapshot 2026-11-21.