<h1>Anthropic’s Claude Code degradation: what happened, why it took seven weeks, and what it means</h1>
<p>For about seven weeks, developers using Claude Code noticed something was off. More errors, shallower responses, instruction-following failures, a tendency to edit code it wasn’t asked to touch. Most assumed it was in their heads. It wasn’t.</p>
<p>On April 20, Anthropic published a postmortem confirming three separate engineering changes had quietly degraded the product — none of which were communicated to users at the time.</p>
<h2>The three changes</h2>
<p><strong>Reasoning effort downgrade (March 4):</strong> The default reasoning effort was quietly dropped from “high” to “medium” to reduce perceived freezing during sessions. Users noticed the model felt dumber. It was restored on April 7 after enough complaints accumulated.</p>
<p><strong>Context bug (March 26):</strong> A bug caused Claude to discard its reasoning history after every single turn, instead of only after an hour of inactivity. The result was a model that kept forgetting what it had already done, leading to repetition and regressions mid-session. Fixed April 10.</p>
<p><strong>Word-count system prompt (April 16):</strong> An overly aggressive system prompt was added to reduce verbosity, capping responses at 25 words in some cases. For a coding assistant, this was a disaster. Reverted April 20.</p>
<p>All three fixes were in place by version 2.1.116. Anthropic also reset usage limits for all subscribers.</p>
<h2>The silence was the real problem</h2>
<p>The fixes matter, but the communication failure is what’s sticking with people. Seven weeks passed between the first change and any public acknowledgment. Users who noticed the degradation had no way to know whether the problem was real, whether it was being worked on, or whether they were just using the tool wrong.</p>
<p>Some developers started posting “conspiracy theories” about what Anthropic had changed. At least one of those theories turned out to be entirely correct — confirmed only after the postmortem dropped.</p>
<p>The argument from engineering leaders is straightforward: a real-time status update during an incident is worth more than a detailed postmortem published weeks later. By the time the postmortem arrived, the speculation had already done its damage.</p>
<h2>The Opus 4.7 situation</h2>
<p>Separate from the Claude Code issues, Anthropic’s latest flagship model — Opus 4.7 — has drawn mixed reactions. It’s genuinely better at following instructions and no longer rewrites code you didn’t ask it to touch. But it also hallucinates more than its predecessor, avoids proactive web searches even when they’re clearly needed, gives shallower responses by default, and burns through up to 35% more tokens due to an updated tokenizer.</p>
<p>The timing didn’t help. Anthropic publicly acknowledged Claude’s quality slip the same week OpenAI launched GPT-5.5. Apologizing for your product while a competitor ships something new is a rough position to be in.</p>
<h2>What this means for building on closed models</h2>
<p>The broader concern here isn’t really about any single bug. It’s about what it means to build on a closed model you don’t control.</p>
<p>In the span of a month, Anthropic silently changed default behavior, introduced a regression that broke session context, added a system prompt that capped response length, banned some corporate customers, and changed plans for users with certain files in their repos — none of it announced in advance.</p>
<p>For individual developers, a week of degraded output is annoying. For teams that have built products or workflows on top of Claude Code, it’s a more serious problem. You can’t debug something you don’t know changed.</p>
<p>The postmortem is a good sign — it’s honest, specific, and takes responsibility. But it came after the fact. The question going forward is whether Anthropic treats this as a one-time cleanup or as evidence that their incident communication process needs to change before the next time something quietly breaks.</p>


Collections

For seven weeks, Claude Code users experienced degraded performance due to three undisclosed engineering changes: a reasoning effort downgrade, a context bug that wiped session history each turn, and an overly aggressive word-count system prompt capping responses at 25 words. Anthropic published a postmortem on April 20 confirming all three issues, with fixes in place by version 2.1.116. The core criticism is the seven-week silence — users had no way to know whether problems were real or being addressed. Separately, the new Opus 4.7 model shows mixed results: better instruction-following but more hallucinations, shallower responses, and 35% higher token usage. The incident raises broader concerns about building products on closed AI models where behavior can change silently without notice.

Claude Code quality regression: Anthropic publishes post-mortem

<p>I can only imagine how hard it is to debug such issues in a non deterministic system</p>
