Anthropic Traces Claude Code Quality Drop to Three Product Changes, Resets Limits

Anthropic said Thursday that three product-layer changes shipped between March and April degraded Claude Code, closing out weeks of user complaints and public pushback from company staff. In an engineering postmortem, the company traced the drop to a March 4 reasoning-effort downgrade, a March 26 caching bug that cleared thinking blocks on every turn instead of once, and an April 16 verbosity instruction that later showed a 3% hit on one coding evaluation. Anthropic said it resolved all three issues by April 20 in v2.1.116 and reset usage limits for every subscriber as of April 23.

Three bugs, overlapping on purpose

On March 4, Anthropic changed Claude Code's default reasoning effort from high to medium to cut tail latencies that occasionally made the UI appear frozen. Internal evals said intelligence would drop only "slightly." Users said otherwise. The company reverted on April 7 and now defaults Opus 4.7 to xhigh and every other model to high.

A separate shipment on March 26 turned out to be the most damaging. Cherny explained on Hacker News that resuming a long-idle session meant a full cache miss, so engineers wanted to prune old thinking blocks from sessions idle for more than an hour, using the clear_thinking_20251015 header with keep:1. The optimization was supposed to run once per stale session. A bug caused it to fire every turn for the rest of the session. Claude kept executing tool calls, but with progressively less memory of why it had made them. Anthropic said it believes this drove separate reports of rate limits draining faster than expected, because each pruned request produced a cache miss. Fixed April 10 in v2.1.101.

The last regression arrived with Opus 4.7 on April 16. To curb the new model's verbosity, Anthropic added one line to the Claude Code system prompt: "Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail." Broader ablation testing during the quality investigation later showed that one line cut coding-eval performance by about 3% on both Opus 4.6 and 4.7. Reverted April 20.

A forensic case from AMD opened the door

A major public flashpoint came on April 2, when Stella Laurenzo, identified in VentureBeat and Business Insider reporting as a senior director in AMD's AI group, filed a GitHub issue arguing that Claude Code could no longer be trusted for complex engineering work. Her evidence was not a vibe check. She pulled 6,852 session files from her team's production logs. Those logs captured 234,760 tool calls and 17,871 thinking blocks. Files read before editing had dropped from 6.6 to 2.0. Median visible thinking length had collapsed from roughly 2,200 characters to about 600.

Claude Code lead Boris Cherny thanked her for the analysis but disputed her main conclusion, pointing to a UI-only change and the earlier medium-effort default. He was blunter on X, reportedly replying "This is false" to accusations that Anthropic had secretly nerfed the model. Three weeks later, Anthropic's postmortem confirmed that a real regression had shipped, though not the specific mechanism Laurenzo identified. Cherny later wrote in a follow-up X thread, per coverage from Kingy.ai, that "this has probably been the most complex investigation we've had" and that "the root causes were not obvious, and there were many confounders."

What engineering review missed

The caching bug slipped past multiple human and automated code reviews, unit tests, end-to-end tests, and internal dogfooding. A separate internal experiment that changed how thinking was displayed suppressed the bug in most CLI sessions Anthropic's own engineers used. And the corner case required a session to cross the one-hour idle threshold once, which rarely happened in short test runs.

When Anthropic back-tested its enterprise Code Review tool against the offending pull requests, Opus 4.7 flagged the defect. Opus 4.6 did not. Anthropic sells AI code review as a coding-assistant feature, and its own pipeline missed a bug that its newer model found in retrospect.

The trust debt

The bigger cost is trust. For most of April, named Anthropic staff publicly disputed claims of deliberate degradation and pointed users to UI changes or effort-setting explanations, while developers kept posting telemetry that suggested a broader product regression. Claude Code is already large enough that even small regressions can compound quickly across a paying customer base. Users felt blindsided, not because bugs happened, but because the official line from Cherny and others had been that the redaction header was UI-only and the evals looked clean, a framing that held right up until the company admitted three separate regressions had shipped.

That pattern repeats elsewhere on the Claude surface this month. On April 21, Anthropic briefly pulled Claude Code from its $20 Pro plan in a small pricing test that Ars Technica reported had landed on roughly 2% of new prosumer signups while the public pricing page was updated for everyone. Head of growth Amol Avasare reversed the page change within a day and apologized for the confusion. The two episodes share the same pattern: communication lags that leave paying users to discover product changes through social media.

What changes next

Every Claude Code system prompt change now runs through a full per-model eval sweep with ablations, gated by new audit tooling, and model-specific instructions must be scoped in CLAUDE.md. Future changes that trade against intelligence, whether for speed or cost, ship with a soak period and a gradual rollout. More Anthropic staff will move onto the exact public build of Claude Code. You can follow the new @ClaudeDevs account on X for the explanation work that has been falling to scattered employee replies.

Users will judge the next regression on two clocks: how fast Anthropic spots it, and how fast it says so in public.

Karpathy Killed His RAG Pipeline for a Folder of Markdown. Here's the Full Build Guide.

The Implicator

AI dubbing is not translation. It is a rights transfer in disguise.

In December 2025, anime fans opened Prime Video and heard something wrong. The words arrived in English and Latin American Spanish. The plot remained legible. But the voices, in titles including Banan

The Implicator

Bollywood's AI Films Score 1.4 on IMDb. Millions Watch Anyway. That's the Point.

A show based on the ancient Hindu epic "Mahabharat" holds a 1.4 out of 10 on IMDb. Reviewers call the lip-sync broken, the styling unnatural, the sequences cheap. The series has logged 26.5 million vi

The Implicator

Harkaram Grewal

New Delhi

Freelance correspondent reporting on the India-U.S.-Europe AI corridor and how AI models, capital, and policy decisions move across borders. Covers enterprise adoption, supply chains, and AI infrastructure deployment. Based in New Delhi.