Anthropic fixes Claude Code performance issues caused by unintended system changes

Anthropic has acknowledged that recent performance issues with its flagship AI tool, Claude Code, stemmed from unintended changes to its internal "harnesses" and operating instructions, not deliberate downgrades as some users speculated. The company identified three specific problems after weeks of complaints from developers and power users reporting worse reasoning, more hallucinations, and inefficient token usage, according to a detailed blog post and reports from VentureBeat and Business Insider.

For several weeks leading up to the announcement, frustration had been building across platforms like GitHub, X, and Reddit. Users described the degradation as "AI shrinkflation," with one AMD senior director noting on GitHub that Claude had regressed to the point where it could no longer be trusted for complex engineering tasks. Complaints intensified following the release of Claude Opus 4.7, where benchmarks showed long-context retrieval dropping sharply from 78% to 32%, and a new tokenizer consuming up to 35% more tokens for the same inputs, as highlighted in user discussions and analyses shared on Axios and YouTube breakdowns.

Anthropic firmly denied accusations of "nerfing" the model—intentionally weakening it to cut costs or redirect compute to projects like the advanced Mythos model. Instead, the issues were product-level tweaks, including adjustments to default reasoning levels, shortened prompt cache times from one hour to five minutes, and changes to system prompts that reduced depth in responses. "We take reports about degradation very seriously. We never intentionally degrade our models," the company stated in its Thursday blog post, as reported by Business Insider.

By April 20, Anthropic had rolled out fixes for all three issues, including tighter controls on system prompt changes, an improved code review tool, and having more staff test the public build of Claude Code. The company also reset usage limits for all subscribers and thanked users for their feedback, which helped validate long-standing concerns. Some users expressed relief, resharing earlier complaints now confirmed by the official response.

This episode underscores the growing dependence on AI coding assistants among developers, where even subtle backend changes can disrupt workflows and erode trust. Power users, in particular, felt the sting as they balanced intelligence with speed in high-stakes engineering tasks. Anthropic's swift response and preventive measures signal a commitment to transparency amid its rapid growth, including a $1 trillion valuation on secondary markets.

Looking ahead, the fixes aim to restore Claude Code's edge, but users will be watching closely for sustained improvements. With competitors vying for dominance in AI tools, incidents like this highlight the challenges of scaling sophisticated models without alienating the technical community that drives adoption. Anthropic's steps to involve more internal testing and monitor changes could set a precedent for the industry.