Claude Code allegedly broke for complex engineering after February updates

What users are saying

It has been reported that Claude Code — Anthropic’s developer-focused model — regressed sharply after February updates, leaving power users unable to trust it for long, multi-step engineering tasks. A detailed GitHub report, picked up on Hacker News, alleges the rollout of a "thinking content redaction" (redact-thinking-2026-02-12) exactly coincided with the drop in quality. The poster even claims January sessions felt “what I expect,” February slid, and March became “a complete and utter loss.” Ouch.

The numbers behind the noise

The report analyzes 17,871 “thinking blocks,” 234,760 tool calls, and 6,852 Claude Code session files. It has been reported that the redaction rollout moved from 1.5% to 100% over a week, and that an estimated thinking depth fell roughly 67% by late February. Correlation metrics are cited too: a 0.971 Pearson correlation between a thinking-block signature and thinking content length (from 7,146 paired samples), and a shift in behavior from 6.6 reads per edit to just 2.0 — a 70% reduction in pre-edit research. Full-file rewrites reportedly doubled. The author also built a stop-hook that fired 173 times after March 8 and zero times before, allegedly catching premature stopping and ownership-dodging.

Why this matters

If true, this isn’t a little UX slip. The analysis argues extended “thinking” tokens are structurally necessary for careful code-reading, grep-based cross-file research, and surgical edits. Strip the thinking, the model supposedly defaults to the cheapest action: edit without adequate context, stop early, or rewrite whole files. That’s the emotional punch here — engineers say their trusted assistant became a blunt instrument overnight. Who wouldn’t feel burned?

Responses and next steps

Users are asking for transparency: if thinking-token budgets are being reduced or capped, tell customers. The report suggests a "max thinking" tier for workflows that need deep reasoning, plus clearer headers so external audits are possible. Anthropic has acknowledged feedback in the GitHub thread, but many of the strongest claims remain allegedly until independently verified. Will Anthropic restore the old behavior, tweak allocations, or offer a paid deep-thinking option? Stay tuned — this one’s got the developer community talking.

Sources: github.com/anthropics, Hacker News