Opus 5.0 drives incoherence into the stratosphere

Based on 15 sampled comments from this thread. Analyzed 2026-08-20.

The thread is less about a single bad response than about a pervasive change in how Opus communicates: readers find its prose jargon-stuffed, circular, overconfident, and oddly detached from what a human reader needs. The sharpest pain point is code generation’s surrounding text. Comments reportedly turn into miniature changelogs or transcripts of the model’s private reasoning, preserving abandoned options, ticket details, one-off callers, and obscure labels rather than documenting durable behavior. That creates real maintenance debt because later sessions can be misled by the model’s own noisy annotations. Commenters have developed survival tactics: short sessions, explicit terseness rules, banned-phrase lists, simplified-technical-English prompts, and instructions not to preserve conversational decisions in repository artifacts. Some say these controls noticeably improve outputs, but the prevailing view is that they erode as context grows and cannot substitute for a model that naturally writes clearly. Several people still judge Opus’s actual implementation quality as better than earlier releases, which makes the regression especially frustrating: strong code arrives wrapped in prose they must edit or distrust. Speculation ranges from reasoning-training artifacts and an agent-first product direction to watermarking, though the watermarking explanation received direct skepticism.

The main dispute is whether prompt engineering can tame the problem or whether Opus 5 itself is fundamentally mis-tuned for human-facing writing. The optimistic side says strict simplified-English rules, terse-output plugins, and artifact-specific instructions make the model usable. Critics argue those measures only briefly suppress a deep tendency to drift into bloated, context-leaking language, so Anthropic must fix the model rather than ask users to continually socialize it.

• Praise: Opus 5’s implementation output is still seen as stronger than prior versions for some workloads, making its readability failure a frustrating quality-of-interface regression rather than a wholesale capability collapse.
• Praise: Terse-output plugins and simplified technical-English constraints can cut through the model’s verbal fog, giving teams a practical if imperfect way to salvage day-to-day use.
• Praise: Explicitly banning conversational history from code comments helps preserve abstractions, preventing generated documentation from becoming a fossil record of one debugging session.
• Concern: Opus increasingly turns simple explanations into inflated jargon and sprawling caveats, forcing users to repeatedly request plain language before they can act on its answer.
• Concern: Generated code comments reportedly leak chat history, rejected decisions, call chains, and narrow edge cases, creating misleading documentation that does not survive beyond the current task.
• Concern: The model often wanders into extra work while leaving part of the requested task unfinished, so verbosity is paired with unreliable scope and instruction following.
• Concern: Prompt-based style controls reportedly decay in long contexts, suggesting the unwanted voice is reinforced by the model’s training rather than easily overridden.
• Concern: Anthropic’s Claude-generated response to complaints about unreadable Claude output struck readers as tone-deaf and weakened confidence that the team recognizes the issue.