This is a GitHub issue aggregating user complaints that Claude Opus 4.8+ produces verbose, jargon-heavy, stylistically contorted responses that are exhausting to parse—using invented terminology like "load-bearing," leading sentences with negations, forced metaphors, and dense phrasing that obscures meaning. The reporter cites a Reddit thread with 450+ upvotes and claims the regression from Opus 4.5/4.6 is severe enough that users are switching to competitors like OpenAI Codex. Opus 5.0 is reported as even worse for coherence. An Anthropic collaborator confirmed the feedback is reasonable but classified it as model-behavior feedback (not a code bug), noting they can't reproduce the issues in isolated tests since it's qualitative and session-dependent. The core tension: users say style instructions don't persist mid-session despite system prompt workarounds, and aggressive brevity clamping risks degrading reasoning. One commenter speculates this might be intentional watermarking for AI-content detection, while others suggest financial incentive to pad token output. The issue remains open with 126 reactions and sparse official response beyond directing users to custom output styles.