V4.1 providers may not be following the intended reasoning-effort mapping
A user reports seeing medium/high reasoning settings mapped to 50 and argues that inference providers should follow the model’s intended mapping.
TLDR
A user described V4.1 as the first model they had seen with a numerical reasoning-effort parameter that directly influences the length penalty. In a September 15 follow-up, they said many inference providers appeared not to follow the intended mapping, citing medium/high settings mapped to 50.
Combined views
3.4K
1 Source, first seen 15d ago