Claude Code regression
- Anthropic said on April 23 that three separate changes degraded Claude Code, Claude Agent SDK, and Claude Cowork over weeks, after users spent much of March and April complaining that coding quality had slipped. - The company said the problems ran from March 4 to April 20 and were fully resolved in version 2.1.116, after reversions on April 7 and April 20 and a bug fix on April 10. - The episode hit a flagship developer product as Anthropic’s revenue run rate reached $30 billion and customers questioned its transparency after weeks of denials. (anthropic.com)
Anthropic said three internal changes, not user error, caused Claude Code to get worse over several weeks in March and April. (anthropic.com) The company said the issues affected Claude Code, the Claude Agent SDK, and Claude Cowork, but not the Claude application programming interface, or API. Anthropic said all three were resolved by April 20 in version 2.1.116. (anthropic.com) The first change landed on March 4, when Anthropic lowered Claude Code’s default reasoning effort from high to medium to cut long delays that sometimes made the interface look frozen. Anthropic said it reversed that change on April 7 after users said they preferred higher intelligence by default. (anthropic.com) The second problem started on March 26, when a fix meant to trim old “thinking” from sessions idle for more than an hour kept firing on every turn instead of once. Anthropic said that bug made Claude seem forgetful and repetitive until it fixed the issue on April 10. (anthropic.com) The third change arrived on April 16, when Anthropic added a prompt instruction to make Claude less verbose. The company said that instruction, combined with other prompt changes, hurt coding quality and was removed on April 20. (anthropic.com) Anthropic said the failures were hard to spot quickly because each change hit a different slice of traffic on a different schedule. The result, it said, looked like broad but inconsistent degradation rather than one clean outage. (anthropic.com) The company said its internal usage data and evaluations did not initially reproduce what users were reporting. As of April 23, Anthropic said it was resetting usage limits for all subscribers. (anthropic.com) The public admission followed weeks of user backlash and cancellations, with some customers saying Anthropic had implied nothing was wrong or that users were misreading the product’s behavior. Fortune reported that the dispute also fed speculation that Anthropic was rationing scarce computing capacity. (aol.com) Anthropic told Fortune that demand for Claude had grown at an “unprecedented rate” and that its infrastructure was strained, especially at peak hours. The company said it was expanding capacity through Amazon and Google partnerships and denied that the Claude Code problems were anything other than the technical issues in its postmortem. (aol.com) Claude Code launched in early 2025 and became a key product for solo developers and enterprise teams. Fortune reported that Anthropic’s annualized revenue run rate reached $30 billion, more than triple the figure at the end of 2025. (aol.com) The company’s account leaves a narrower conclusion than many users feared: Anthropic says Claude Code was not secretly downgraded across the board, but it was degraded by a stack of shipping decisions and one bug. Anthropic’s fix was to revert two changes, patch the third, and promise tighter checks before similar updates go live. (anthropic.com) (the-decoder.com))
