Anthropic confirms 6-week Claude regression
- Anthropic said on April 23 that Claude Code really did get worse for weeks, and pinned the drop on three separate product changes. - The timeline matters: one change landed March 4, another March 26, a third April 16, and the full fix only arrived April 20. - It matters because the model itself was not degraded — the break happened in defaults, caching, and prompts around it.
Claude Code really did regress for about six weeks. Anthropic said so in a public postmortem on April 23, after a month of users saying the tool felt dumber, more forgetful, and strangely inconsistent. The important part is that this was not a secret model downgrade. Basically, the model weights and the API were fine. The failure happened in the product layer wrapped around the model — defaults, session handling, and prompt instructions. (anthropic.com) ### What actually broke? Three separate changes stacked on top of each other. On March 4, Anthropic changed Claude Code’s default reasoning effort from high to medium. On March 26, it shipped a session change meant to clear older thinking after more than an hour of idle time, but a bug kept clearing that history every tu(anthropic.com) responses less verbose, and that ended up hurting coding quality. (anthropic.com) ### Why did lowering “reasoning effort” matter? Because in Claude Code, more effort usually means better answers — but slower ones. Anthropic says it made the March 4 change to reduce long latency spikes that made the UI look frozen for some users. The tradeoff turned out to be the wrong one. Users would rather wait a bi(anthropic.com)ult, so Anthropic reverted that change on April 7. (anthropic.com) ### What was the caching bug doing? This is the sneaky one. The intended behavior was simple: if you came back to an old session after more than an hour, Claude would clear older thinking once so the session could resume faster. But the bug made that cleanup happen over and over on every turn for the rest of the session. (anthropic.com) — exactly the kind of thing developers had been complaining about. Anthropic says it fixed that on April 10. (anthropic.com) ### Why would “less verbose” make coding worse? Because shorter is not the same as better. Anthropic added a prompt instruction on April 16 to reduce verbosity, but in combination with other prompt changes it pushed answers too far toward brevity. For coding, that can mean less explanation, weaker planning, and thinner step-by-step reasoning. Anthropic reverted that prompt change on April 20. (anthropic.com) ### Which products were hit? Not everything. Anthropic says the API and inference layer were unaffected. The problems hit Claude Code, the Claude Agent SDK, and Claude Cowork. The reasoning-effort and caching issues affected Sonnet 4.6 and Opus 4.6. The verbosity issue reached Sonnet 4.6, Opus 4.6, and Opus 4.7. (anthro([anthropic.com)y did this feel so hard to diagnose? Because the failures were staggered. Different changes hit different slices of traffic on different dates, so users saw a messy mix of symptoms instead of one clean outage. Anthropic also says its internal usage and evals did not initially reproduce the issues, which is a pretty(anthropic.com)nd of degradation real developers felt in live workflows. (anthropic.com) ### What did Anthropic do after fixing it? All three issues were resolved by April 20 in Claude Code v2.1.116, and Anthropic said on April 23 that it was resetting usage limits for all subscribers. That matters because users were not just getting worse answers — some were also burning through limits while the tool spent tokens on unhelpful behavior. (anthropic.com) ### What’s the bigger lesson here? AI products are not just “the model.” They are a stack. Change the defaults, the memory behavior, or the system prompts, and you can make a strong model feel suddenly mediocre without touching the core model at all. That is the real story here — and it is why postmortems like this matter. (anthropic.com)
