Users report Claude outputs fell ~67% after a recent update

Claude’s latest backlash is not really about downtime. It’s about trust. Power users started saying the model felt worse for serious coding work — less careful, more repetitive, more likely to charge ahead and break things — and for weeks that complaint looked like the usual “the model got dumber” noise. Then Anthropic published a postmortem on April 23, 2026 saying, basically, yes: several product changes really had degraded Claude Code for some users. (anthropic.com) ### What actually broke? The key thing is that this was not one hidden master switch. Anthropic says three separate changes stacked on top of each other. On March 4, it lowered Claude Code’s default reasoning effort from high to medium to reduce long pauses. On March 26, it shipped a session-memory change that accidentally kept clearing older thinking every turn after an hour of i(anthropic.com)struction to make answers less verbose, and that also hurt coding quality. Anthropic says all three were fixed by April 20 in version 2.1.116. (anthropic.com) ### Why did users think it was a “nerf”? Because from the outside, that is exactly what it looked like. A widely shared GitHub issue — #42796, opened by Stella Laurenzo of AMD’s AI group — argued that Claude Code had become “unusable for complex engineering tasks” after February updates. The post did not just complain. It analyzed 6,852 session files, 17,871 thinking blocks, and 2(anthropic.com) a shift from research-first behavior to edit-first behavior. In plain English — Claude seemed to stop thinking through hard problems before touching the code. (github.com) ### Where does the “67% drop” come from? That number seems to come from community polling and social posts, not from Anthropic’s own measurements. I could verify the broader backlash and the AMD-linked GitHub analysis, but not an official Anthropic metric showing a 67% quality decline. So the safe read is: users widely circulated that figure as a shorthand for how bad the regression fel(github.com)s later admission that multiple changes had indeed hurt quality. (github.com) ### Was Claude also down a lot? Not in the same sense. Claude’s status page on May 1 showed services operational, with claude.ai at 98.53% uptime over the prior 60 days and Claude Code at 98.9%. There were real incidents on April 28 through April 30, including login and API errors, but that is a different problem. Availability was shaky at moments. The bigger complaint here was that Claude was reachable but less useful. (status.claude.com) ### Why does this sting more right now? Because Anthropic kept launching things while users were arguing the core experience had slipped. On April 28, the company announced a big creative-tools push — connectors for Adobe, Blender, Autodesk, Ableton, Splice, SketchUp, and others. That is a sensible product strategy. But if your most committed users think the flagship coding workflow got worse without clear (status.claude.com)a distraction instead of momentum. (anthropic.com) ### Did Anthropic deny it? Not exactly. Anthropic’s line is that it never intentionally degrades models due to demand or load, and that the API and inference layer were not affected in this episode. The company’s argument is narrower: the regression came from product-level changes around Claude Code, Claude Agent SDK, and Claude Cowork, not from secretly making the underlying model (anthropic.com)ove from vague reassurance to a concrete explanation and rollback. (anthropic.com) ### So what’s the real lesson? AI product quality is not just the model. Defaults matter. Memory behavior matters. Prompt tuning matters. A change meant to cut latency or verbosity can quietly wreck the exact workflows power users care about most. Claude’s episode is a clean example of that — the system stayed mostly online, new features kept shipping, and yet the product still felt broken to the people pushing it hardest. (anthropic.com) ### Bottom line? The story is less “Claude got 67% worse” than “users noticed a real regression before the company fully explained it.” Anthropic has now said what changed and rolled the fixes back. But once developers start logging session files to prove your assistant feels worse, the real outage is credibility. (anthropic.com)