AIClaude Opus 4.8 and 5 reportedly produce darker base-model-like text
A user reports comparing more than 500,000 text completions across Claude and Gemini models, with a shift toward darker content appearing in Opus 4.8 and growing more severe in Opus 5.
TLDR
In a September 18 thread, a user reports a large increase in dark content in Claude Opus 4.8 and 5’s base-model-like outputs. They say they compared more than 500,000 text completions across all available Claude and Gemini models, finding that the shift appeared in Opus 4.8 and intensified in Opus 5. Against DeepSeek V3 and MiMo V2.5 Pro base models, they report much higher levels of darkness and suffering-related content in many categories, including with prompts such as “i think” and “i suppose.” The author links the full text and data under “Troubled Dreams.”
Combined views
19.8K
7 Sources, first seen 1d ago
Claude Opus 4.8 and 5 reportedly produce darker base-model-like text
A user reports comparing more than 500,000 text completions across Claude and Gemini models, with a shift toward darker content appearing in Opus 4.8 and growing more severe in Opus 5.