Experiment
One missing connection
updated
September update: Astra and Fable 5.1 on a much larger diagram.
I made a test to understand the current capabilities of SOTA LLMs in visual content understanding, including the recently published DeepSeek OCR with a bunch of people claiming that it solves the context length limitation, memorization, and maybe your back pain too.
The prompt was simple: “What’s the difference between these two diagrams?”
Show the difference
The arrow from Node 6 to Node 8 appears in A but is missing in B. Node 6 still connects to Node 7 in both.Results
The largest models have mixed results (one-shot):
| Model | Response |
|---|---|
| GPT-5 | Found the missing connection. |
| Sonnet | Said the diagrams were identical. |
| Gemini | Said the diagrams were identical. |
| Qwen | Invented a missing connection elsewhere. |
DeepSeek OCR has also failed the task, although I had to stitch the images because I couldn’t input two of them at the same time.
Check it out for yourself (tested on 12GB 3060).
Update · 13 September 2026
I kept benchmarking every new model release on a much larger pair of screenshots of the CAP system diagram, with one redacted edge. Astra and Fable 5.1 are the first ones to notice the difference.

