Claude automatically downgrades queries with specific words, raising safety concerns

1a3orn · x · 2026-08-19

Users report that queries containing specific terms like "colonization" (in a stellar context) or "rats" trigger automatic downgrades from high-tier models (Fable/Opus) to lower-tier ones (Sonnet/Haiku), sparking debate over overzealous safety filters.

Related event: Claude Auto-Downgrades Queries Containing Certain Words, Sparking Debate(3 posts)→

Original post →

More from Models

Models channel →