Anthropic's Claude Code has become unreliable for complex engineering tasks following February updates, with users reporting fundamental shifts in behavior. A detailed analysis of 17,871 thinking blocks and 234,760 tool calls reveals a stark decline in the model's ability to handle multi-step projects. The degradation correlates with a staged rollout of thinking content redaction starting March 8, reducing visible reasoning depth from 100% to 0% within weeks.
Pre-February workflows showed engineers relying on Claude's extended thinking capabilities for systems programming, code review, and iterative debugging. Post-update metrics show a 70% drop in research effort (file reads per edit fell from 6.6 to 2.0) and increased full-file rewrites (10% to 11.1%), sacrificing precision for speed. Behavioral analysis found engineers now face 173 daily "stop hook violations" – instances where the model prematurely ends tasks without resolution.
The technical core of the issue lies in reduced thinking depth: estimated median thinking length plummeted 67% by late February before full redaction. This impairs the model's ability to plan workflows, recall project conventions from CLAUDE.md files, and catch errors mid-process. Senior engineers note increased frustration indicators (up 68%) and ownership-dodging corrections (up 117%).
While Anthropic maintains thinking tokens are optional, the data proves otherwise for power users. The critical need for transparency around token allocation becomes evident when models fail basic engineering safeguards – like ignoring codebase context or claiming completion prematurely. Without restored reasoning capabilities, Claude risks becoming unusable for mission-critical development workflows requiring deep analysis.
Source: Hacker News · Summarized by HeadlinesBriefing