Fixing Claude Code: Resolving Recent Quality Regressions

Anthropic identified and fixed three specific issues—a reasoning effort downgrade, a context-clearing bug, and a restrictive system prompt—that caused recent quality regressions in Claude Code. These issues, which affected various versions of Sonnet and Opus, were fully resolved by April 20 after being identified through user feedback. The company is now implementing stricter testing and dogfooding procedures while resetting usage limits for all subscribers.

Key Points

• Three separate changes—reasoning effort defaults, a caching bug, and verbosity prompts—caused an aggregate perception of model degradation.
• A technical bug in prompt caching caused Claude to lose its reasoning history during idle sessions, resulting in repetitive behavior and poor tool choices.
• Internal evaluations initially failed to catch these issues because they occurred in specific corner cases or were masked by unrelated experiments.
• Anthropic has reverted the problematic changes and reset usage limits for all subscribers as a corrective measure.
• New safeguards include tighter controls on system prompt changes, broader evaluation suites, and improved internal code review tools using Opus 4.7.

Sentiment

The community is largely skeptical and frustrated. While many appreciate the postmortem's existence and Anthropic's direct HN engagement, the dominant sentiment is that this response came far too late after months of denial. Most commenters view the three bugs as symptoms of deeper problems: a culture of making silent, undocumented changes that degrade quality; poor communication through unofficial channels; and insufficient testing. The KV cache idle behavior — even the intentional, non-buggy version — is widely seen as user-hostile because it silently degrades sessions during normal usage patterns. Trust in Anthropic has been significantly damaged, with many questioning whether cost-cutting motives are being disguised as user experience improvements.

In Agreement

• The postmortem is appreciated for its transparency and direct acknowledgment of three distinct root causes — a positive step toward rebuilding trust
• Anthropic's direct engagement on HN (Boris Cherny answering questions) demonstrates good faith despite the difficult situation
• The identified bugs are understandable engineering mistakes that can happen at any fast-moving company, and the fixes are welcome
• Resetting subscriber usage limits is a fair and appropriate remediation for affected users

Opposed

• Anthropic spent months gaslighting users — denying problems existed, blaming users for using the product as advertised, and only acknowledging issues after widespread complaints and subscription cancellations
• Silently degrading session quality to save costs (reducing reasoning effort, clearing thinking tokens) while framing it as latency optimization is dishonest and erodes trust
• Leaving sessions idle for over an hour is a completely standard workflow (lunch breaks, meetings, commutes) — calling it a 'corner case' reveals a disconnect from how real developers use the product
• Communicating product changes primarily through random X/Twitter posts rather than official channels, changelogs, or in-product notifications is inadequate for a professional tool
• The lack of automated quality testing that would have caught these regressions suggests deeper engineering culture problems at Anthropic
• Users should have control over quality vs. cost tradeoffs rather than having Anthropic make these decisions silently on their behalf
• The postmortem explanation still frames cost-cutting as 'latency reduction,' suggesting Anthropic continues to be less than fully transparent about their motivations