Opus 5.0 drives incoherence into the stratosphere

Anthropic’s latest Claude Opus 5 model is drawing sharp criticism from heavy users who say its prose has become bloated, jargon-filled, and increasingly incoherent, with recurring “Claudisms” like “load‑bearing” seeping into real-world codebases, comments, and documentation. Many report that attempts to constrain its tone or verbosity quickly fail, forcing them to add linters, banned-word lists, or Simplified Technical English prompts, or to switch to alternative models for day-to-day work. Speculation ranges from alignment and agent-oriented training to watermarking and synthetic data as causes, amid growing concern that Anthropic is optimizing more for autonomous agents and internal tooling than for human-readable output.

Perceived regression in Opus 5.0

  • Many users see Opus 5 (and late 4.x) as a downgrade in day‑to‑day usability, especially for coding.
  • Complaints center on incoherent, hard‑to‑read prose and planning summaries, despite often good code quality.
  • Several people say they’re switching or trialing alternatives (GPT 5.6 Sol, Codex, sometimes Gemini) and even downgrading paid plans.

Language style complaints

  • Frequent, idiosyncratic “Claudisms” (e.g., “load‑bearing”, obscure terms, dense abstractions) now saturate codebases, comments, PRs, and docs.
  • Output is described as verbose, jargon‑heavy, over‑hedged, and oddly metaphorical (“agent speak”), with comments that narrate history and side‑quests instead of describing current code.
  • Some find this merely annoying; others describe genuine stress and burnout, calling the style “toxic” or unreadable.

Reasoning and agent behavior

  • Users report poor “reading the room”: unexplained internal labels (“disposition 7”), numbered plans, and self‑referential reasoning leaking into final text.
  • Models often resist simple instructions, over‑scope tasks, or second‑guess user intent, especially in agentic/code flows.
  • Some say reasoning quality itself has degraded; others think the underlying competence is high but buried under bad communication and planning artifacts.

Speculated causes

  • Hypotheses include watermarking pressure, chain‑of‑thought/agent tuning leaking into user‑visible text, synthetic data/model collapse, and optimization for autonomous agents over human readability.
  • A few argue watermarking alone can’t explain the magnitude of the style shift; cause remains unclear.

Mitigation strategies

  • Common tactics: “ELI5/ELI10” prompts, “explain like to a colleague,” Simplified Technical English (ASD‑STE100), banned‑word lists, terseness plugins, and custom styleguides.
  • Many say such steering works only for 1–2 turns before the model “drifts back.”
  • Some separate planning vs. implementation sessions, forbid AI‑authored commits, and manually clean comments.

Impact on teams and workflows

  • Teams add linters/hooks to block certain phrases or comment patterns; some fail builds on Claudisms.
  • Extra tokens and time are spent cleaning AI prose, undermining productivity gains.

Reactions to Anthropic’s handling

  • Several see AI‑generated official replies to these complaints as tone‑deaf or disrespectful.
  • There is broader unease about organizational “AI worship” and the risk of AI systems subtly shaping company culture and hiring.