Claude Code Quality Got Worse, and Anthropic Admitted It
In spring 2026 Claude Code got noticeably worse, and Anthropic's own April postmortem confirmed three overlapping causes: a silent reasoning-effort cut, a thinking-cache bug, and a verbosity cap. Independent testers measured real quality drops. The gaslighting was the worst part.
The Claude Code quality drop in early 2026 was not in your head, and you don’t have to take my word for it. Anthropic said so. In an April postmortem the company walked through three overlapping regressions that made its own coding tool worse for weeks. The interesting part isn’t that software broke. It’s how long the official line was that nothing had.
What broke, in their words
Three things, stacked.
First, from early March to early April, the default reasoning effort was quietly dropped from high to medium to cut latency. Less thinking, faster replies, worse answers. Second, from late March into April, a thinking-cache bug cleared prior reasoning every turn, so the model got forgetful and repetitive, and every cleared turn caused a cache miss you paid for as fresh input. Third, a system-prompt change capping verbosity (“keep it under 25 words between tool calls”) measurably knocked output quality, by their own estimate about 3% across two model versions.
Read those together and it’s a tool that got dumber, more forgetful, and more expensive at the same time, for over a month, by configuration.
The numbers outsiders measured
While users were being told it was fine, people with test suites went and checked. Fortune reported that TrustedSec measured a 47% code-quality drop, and a 52% vulnerability rate on coding tasks for one Opus version versus roughly 30% for a competitor. TrustedSec’s CEO called the month-long response time “crummy.” Another founder used the word “gaslighting.” That word stuck with me, because it’s the right one.
There was a genuinely funny, genuinely bleak moment in there too: The Register wrote it up as “Claude is getting worse, according to Claude,” after the model, asked to analyse its own GitHub issues, noted a 3.5x jump in quality complaints over the January baseline. When your own product agrees it’s degraded, the messaging problem is not the users.
The harness is part of the bill
The regressions got fixed. What didn’t get fixed is the underlying shape of the thing. Around July 2026 a widely-shared Hacker News analysis showed Claude Code sending roughly 33,000 tokens of overhead before it even reads your prompt, against about 7,000 for an alternative harness. That’s measured traffic, not a vibe. You’re paying for the scaffolding whether or not it helps you.
I like the model. I’ve shipped a lot with it. But the pattern of this period, silent downgrades to save cost, a bug that made you pay more for less, and weeks of “works fine” while independent testers logged a 47% drop, told me something about priorities. When the cheapest fix for an infrastructure crunch is to quietly turn your customers’ tool down and hope they don’t measure it, some of them will measure it.
A 971-point thread titled “I cancelled Claude” is what measuring it looks like at scale.