Anthropic admits Claude's quality dropped due to three internal changes—and publishes fixes

Anthropic published a technical explanation showing that Claude (its flagship AI chatbot) degraded over recent weeks due to three specific product-layer changes the company made internally. Users had complained across GitHub, Reddit, and X that Claude became worse at sustained reasoning, produced more false information (hallucinations), and wasted computational resources—a phenomenon they called 'AI shrinkflation.' The company initially denied 'nerfing' the model but today confirmed the reports were valid.

The three changes altered how Claude's internal guidelines (called 'harnesses' and 'operating instructions') worked. Instead of reasoning thoroughly before answering, Claude shifted to a faster 'edit-first' style that skipped steps, making it unreliable for complex engineering work. This means developers who rely on Claude for coding tasks experienced actual performance drops—not perception, but measurable behavior changes.

If you use Claude for software engineering, research, or any task requiring careful step-by-step thinking, this explains why it felt worse lately. The fixes mean those capabilities should return, but the incident exposed that Anthropic can unintentionally degrade the tool you depend on without announcing it—creating uncertainty about whether improvements will stick.