Anthropic explains why Claude Code got worse

If you felt like Claude Code suddenly got dumber over the past month, you were not imagining things.

In an unusually transparent postmortem published on April 23, Anthropic confirmed that Claude Code experienced widespread quality degradation. The underlying model weights never changed, nor was the core API impacted. Instead, a series of three separate product and prompt optimizations in Claude Code, Claude Agent SDK, and Claude Cowork cascaded into a mess. Anthropic has resolved all three issues as of April 20 in v2.1.116, and is resetting usage limits for all subscribers.

Three tweaks that broke the agent

The trouble started on March 4. In an effort to reduce UI-freezing tail latencies on Opus 4.6, Anthropic quietly dialed down the default reasoning effort from high to medium. While it trimmed wait times, it noticeably degraded output quality. Even after adding startup dialogs and inline selectors, most users never changed the setting. On April 7, Anthropic reversed course: Opus 4.7 now defaults to xhigh, while other models default to high.

The second issue was a nasty prompt caching bug rolled out on March 26. To cut costs on resuming sessions idle for over an hour, Claude was supposed to purge older thinking tokens just once. Instead, a bug caused it to purge prior reasoning on every single turn for the remainder of the session. Claude effectively lost its working memory, repeating actions and making bizarre tool calls. Worse, every subsequent call missed the prompt cache, quickly burning through subscribers' usage limits. The team fixed this on April 10 in v2.1.101.

The final straw came on April 16. To rein in Opus 4.7's known verbosity, Anthropic added a system prompt restriction: keep text between tool calls to 25 words or fewer, and final responses to under 100 words. That tiny throttle triggered an immediate 3% drop in coding evaluations across both Opus 4.6 and 4.7. Anthropic reverted the prompt on April 20.

Why it matters

This postmortem confirms a long-running suspicion among developers: when top-tier AI models suddenly seem dumber, it is rarely the base model at fault, but rather the fragile scaffolding wrapped around it. In attempting to optimize latency, cost, and token counts, Anthropic inadvertently choked the reasoning capacity that made Claude Code exceptional in the first place.

For paying developers, this means Claude Code is finally back to its sharpest self. If you abandoned it in March or April due to weird context amnesia, updating to v2.1.116 and taking advantage of the reset usage limits makes it well worth another spin.

We think Anthropic handled the aftermath with admirable transparency, laying bare internal engineering blunders that most competitors would sweep under the rug. Still, it highlights a glaring vulnerability across modern AI tooling: as agents depend heavily on multi-turn context and delicate prompt constraints, an innocent 25-word limit can accidentally shave points off model intelligence. Building wrappers around frontier models is proving just as perilous as training them.