Visual Cipher Bench - Spatial: leaderboard

Visual Cipher Bench v1 (2611.16861v1): 40 image-based puzzles, ten each of spatial, logic, rule and knowledge types. Harbor v0.24.0; GPT models use Codex CLI v0.163.0 and other models Claude Code v2.1.296; 60-minute network-disabled runs, 2 CPUs/4 GB/no GPU, Python/NumPy/Pillow tools and read-only puzzle images. Three independent runs per exact model/effort. Accuracy is final-answer exact match after fixed normalization; invalid submissions/timeouts score zero. Each category has 30 attempts; overall has 120. Literal none effort is retained without inferring disabled reasoning. Mean solved-count column is not another benchmark. This board reports spatial exact-answer accuracy in percent; higher is better.

Metric: Exact final-answer accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-6 AstraMax80#6
2Claude Opus 5.5xHigh70#5
3GPT-6.1 SolMax60#10
4GPT-6 LunaMax3.3#99
5Claude Sonnet 5.5Max3.3#19
6DeepSeek V4.1 FlashMax0#62
7Kimi K3.1 ProMax0#29
8Qwen 3.8 MaxxHigh0#59

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=visual-cipher-bench-spatial · How It Works · Data refreshed daily, snapshot 2026-11-16.