Is it just me, or has Claude Opus gotten worse recently?

What happened

Let me check the template file first. Good. The template confirms the news type structure I need to follow. Here is the article:

A growing number of developers using Claude Opus for production coding work are reporting a noticeable drop in reliability since recent updates rolled out. The complaints are specific and consistent: a model that once handled complex, multi-step coding tasks in a single prompt is now failing those same tasks outright, ignoring project-level configuration files, and rewriting code it was never asked to touch.

This piece is for engineers who use Claude Opus as a daily coding assistant — particularly anyone who has wired it into automated workflows, relies on CLAUDE.md project rules to constrain its behavior, or routes it through tools like Codex or Fable. If you have started second-guessing outputs that used to be dependable, you are not alone, and the pattern is worth understanding.

How it works

Reports surfaced on Hacker News from developers who noticed Claude Opus regressing on tasks that previously required no iteration. The core complaint is not that the model is occasionally wrong — it is that behaviors which were once stable appear to have become unreliable in a coordinated way. Developers describe prompts that used to resolve cleanly in one pass now failing completely rather than partially. The affected use case is coding workflows: reading a codebase, respecting constraints defined in project configuration files, making targeted edits, and updating documentation to reflect those edits. Each of those steps is breaking for a subset of users.

The discussion threads pulling these reports together span Claude, Codex, and Fable integrations, which suggests the observed behavior is tied to Claude Opus itself rather than to any one client or plugin layer. No official change log or model version bump has been cited to explain the shift, which makes systematic debugging harder for teams trying to pin down when their workflows started misbehaving.

Claude Opus is a large language model designed for complex reasoning tasks. When developers embed it into coding pipelines, they typically provide a CLAUDE.md file — a plain-text configuration document that encodes project rules, style conventions, and off-limits operations. The model is expected to read and honor that file at the start of every session. Reported behavior suggests the model is now bypassing those rules, producing edits that contradict explicit constraints.

Why it matters

Beyond CLAUDE.md compliance, the failure modes being described point to a breakdown in what could be called verification behavior. A well-functioning coding assistant is expected to read relevant files before editing them, confirm what existing code does before changing it, and cross-check documentation against the code it describes. What developers are reporting instead is the model making edits based on stale inline comments rather than current code, skipping file reads before making changes, and producing documentation that no longer matches the actual codebase — exactly the class of errors that accumulate silently and are expensive to catch after the fact.

For individual developers, a less reliable Opus means more review time, more prompt iteration, and more cognitive overhead on tasks the model used to absorb entirely. That cost compounds quickly in teams that have built autonomous or semi-autonomous coding loops where a human reviews output rather than writing every line. If the model starts ignoring mandatory project rules and touching files outside its stated scope, those loops can produce incorrect merges, broken tests, or silent regressions that take time to trace back to AI-introduced changes.

Who is affected

The deeper concern is trust calibration. Developers who have tuned their workflows around Opus's previous behavior may not immediately recognize that the model has changed — they may assume the fault lies in their own prompts, configuration, or environment. That misattribution wastes debugging time and can lead teams to incorrectly deprecate working pipeline designs in favor of workarounds that address the wrong root cause.

The most directly affected group is engineers running Claude Opus inside coding automation pipelines — tools like Codex and Fable that orchestrate multi-step agentic tasks rather than single-turn chat. These users depend on consistent rule-following and scoped file access because the model is operating with limited human checkpoints. A model that makes unsolicited edits to unrelated files or ignores CLAUDE.md constraints is a higher-risk actor in an automated context than in a supervised one.

Developers who use Claude Opus interactively but for complex, long-context tasks are also affected. Anyone who relies on the model to reason across a large codebase, track which files it has already read, and produce documentation updates aligned to code changes will encounter the same class of failures: guesswork edits, stale-comment reasoning, and documentation drift. Less affected, for now, are users running simpler single-turn prompts where the failure modes — ignoring project rules, touching out-of-scope files — have fewer consequences.

What to watch next

The first thing to verify is whether CLAUDE.md is being read at all in your current sessions. A straightforward test is to place a clearly unique, low-ambiguity instruction in that file and confirm whether Opus honors it on the first prompt of a session before any other context is loaded. If it does not, the issue is at the configuration ingestion layer. If it does for simple instructions but not complex ones, the failure is more likely a context-window or instruction-priority problem that appears under load.

Anthropic has not issued a public statement addressing the regression reports as of this writing. Developers reporting the issue should document specific before-and-after prompt examples, note which tool or API version they are using, and submit them through official support channels to help Anthropic isolate whether the behavior is a deliberate model trade-off, an unintended consequence of a recent update, or an interaction effect with specific client implementations. Watching Anthropic's model release notes and the Hacker News threads on Claude, Codex, and Fable for corroborating reports or an official response is the most productive near-term posture.