The internet is littered with complaints from users over “nerfed” models with degraded performance, but every once in a while, these complaints have some basis.
Anthropic has published a detailed post-mortem acknowledging that Claude Code users did in fact experience real degradation over the past month — and that it was caused by three separate changes that stacked on top of each other in ways that were hard to diagnose. All three issues have been resolved, and the company is resetting usage limits for all subscribers.
Three Bugs, One Ugly Month
The first issue was a deliberate but misjudged tradeoff. On March 4, Anthropic quietly switched Claude Code’s default reasoning effort from high to medium, citing complaints about excessive latency and token consumption in high-effort mode. Users noticed immediately — the model felt less capable. Anthropic reverted the change on April 7, and has since reset defaults to xhigh for Opus 4.7 and high for all other models.
The second issue was a genuine bug. On March 26, Anthropic shipped a caching optimization meant to reduce costs when resuming idle sessions. The logic was sound: if a session had been inactive for over an hour, clear old reasoning blocks once so Claude doesn’t send stale thinking tokens on cache-miss requests. But the implementation had a flaw — it cleared thinking history on every subsequent turn for the rest of the session, not just once. The result was Claude progressively losing memory of why it had made prior decisions: wrong tool calls, repetitive behavior, and apparent forgetfulness. Because the clearing also produced continuous cache misses, it also explains reports of usage limits draining faster than expected. This was fixed on April 10.
The third issue surfaced on April 16, just as Opus 4.7 — which tends toward verbosity compared to its predecessor — was released. Anthropic added a system prompt instruction capping responses to 25 words between tool calls and 100 words for final answers. Their internal evals showed no regressions. Broader ablation testing later revealed a 3% coding quality drop for both Opus 4.6 and 4.7. The line was reverted on April 20.
Why It Took So Long to Find
Because each change hit a different slice of traffic on a different timeline, the combined effect looked like broad, inconsistent model degradation rather than three discrete bugs. None of the issues were caught by Anthropic’s internal evals or unit tests before shipping — a particularly ironic outcome given that Code Review is one of Claude Code’s flagship enterprise features.
Anthropic ran a post-hoc test using Opus 4.7 to back-review the pull requests that introduced the caching bug. It found the defect; Opus 4.6 did not. This has prompted the team to expand repository context available to Code Review for internal reviews — a meaningful admission that the tool they sell needed to be used more rigorously internally.
There is also an obvious irony in the broader context: Claude Code is now effectively 100% written by Claude Code. A product that writes itself, reviewed by AI, shipped with bugs that AI helped catch only in retrospect.
What Anthropic Is Changing
Going forward, Anthropic says more internal staff will use the public build of Claude Code rather than internal test versions. Every system prompt change will require a full per-model eval sweep with ablations, supported by new tooling to make prompt changes easier to audit; model-specific changes will be gated to the models they target. Any change that trades off intelligence for another metric — latency, verbosity, cost — will require soak periods and gradual rollouts.
These are reasonable safeguards, though the caching bug’s ability to survive multiple human and automated code reviews, unit tests, end-to-end tests, and internal dogfooding before hitting production suggests the surface area for failure is larger than process checklists can fully cover.
The Stakes
Claude Code is not a niche product. Its annualized revenue crossed $2.5 billion in February 2026 and accounts for roughly 4% of all public GitHub commits. At that scale, even subtle degradation compounds across millions of developer sessions. Anthropic’s transparency here is notably more detailed than most companies offer post-incident, but the month-long lag between user reports and confirmed root causes is the more important number to watch as the company grows.