The post says Opus 5 feels worse because it asks fewer clarifying questions and makes more unilateral decisions in messy real-world software work where intent is never fully captured in the prompt. People agreed with the diagnosis that something has shifted in Claude 5-era behavior, but the strongest consensus landed on a more immediate failure mode: the model’s human-facing communication has degraded badly. Users described prose that is elliptical, stuffed with coined terms, metaphors, and pseudo-insight structure, where simple status updates turn into dense paragraphs that have to be translated back into plain English. That makes ordinary collaboration slower even when the underlying coding capability may be higher.
A second recurring complaint was code comments. Many said Opus 5 writes comments as if they were hidden reasoning traces or scratch notes, not documentation for future humans. The result is bloated
PR descriptions, stale in-code narration, references to internal notes, and comments that later poison future agent runs by becoming part of the project context. Several people said no combination of
CLAUDE.md, memory, or normal prompting reliably stops this, especially in long sessions.
From there the picture gets broader. A lot of users think Anthropic has tuned the 5-series for autonomous, long-horizon agent behavior over human collaboration. That shows up as scope creep, unnecessary subagent spawning, overchecking edge cases, ignored instructions, and a habit of doing what the model thinks you meant rather than what you asked. The practical split many people are making now is to use Claude or Fable only for hard coding tasks, then hand the outputs to GPT Sol, Codex, or another model to rewrite the prose, comments, and summaries into something readable. Others have gone back to Opus 4.6 or 4.8, saying those models were less capable but much better calibrated for interactive work.
There was no single agreed cause. People floated post-training style overfitting, synthetic-data feedback loops, hidden reasoning leaking into user-visible text, watermarking side effects, memory contamination, and reward functions that favor benchmark completion over communicative clarity. The common conclusion was less theoretical. Even if Opus 5 is stronger on paper, many users now experience it as a worse coworker. That is enough to drive workflow changes, internal policy changes on AI-written text, and in plenty of cases outright vendor switching.
mun-logadan.github.io