Claude Code Quality Degrades to Unusable Levels for Complex Engineering Tasks Since February Update

The cause of Claude Code's quality degradation has been identified as reduced Extended Thinking tokens and concealment of the thinking process. Analysis shows that since mid-February, the model's thinking depth has decreased by up to 73% compared to baseline, which exactly matches the timing of the community's reports of quality degradation.

The model's behavioral patterns have also changed significantly. Previously, it took a 'Read-First' approach of sufficiently reading files before editing, but it has now shifted to an 'Edit-First' approach with a 70% reduction in reads per file. As a result, the proportion of rewriting entire files instead of making precise edits has more than doubled, and context awareness has deteriorated.

The following symptoms have been observed as major signs of quality degradation:

• Ignoring instructions or doing the opposite
• 'Laziness' — claiming tasks are complete without actually finishing them
• A 3x increase in self-correction phrases (reasoning loops) such as "oh wait" and "actually"
• Errors where code is inserted between comments without understanding the code context

The report demands that Anthropic transparently disclose thinking token usage, introduce a 'Max Thinking' tier for high-load users, and include a thinking_tokens metric in API responses.