When designing partition plates from caliper videos, Claude Opus 5.5 misread once like the real object, while GPT-6.1 Sol was correct all 3 times
#Anthropic #OpenAI #Claude Code
When remaking pen holder partition plates with a 3D printer, a design test using caliper photos and videos was conducted. GPT-6.1 Sol (reasoning high) was correct all 3 times, while Claude Opus 5.5 (effort high) misread once, same as the real object.
An attempt was made to remake pen holder partition plates using a 3D printer. When Claude Code (Claude Opus 5.5) was given photos and videos measured with calipers to design the plates, the finished plates did not fit in the pen holder.
Consequently, the same photos and videos were used to add problems to a personal LLM benchmark, OreOre-Bench. In the results, GPT-6.1 Sol (reasoning high) produced plates of a dimension that fit all 3 times. Meanwhile, Claude Opus 5.5 (effort high) made the same misreading as when it failed with the real object 1 out of 3 times, and it did not fit.
This item is a summary. The full, checkable version is at the original source:
Read the original at Zenn (LLM) ↗