AMD's senior director of AI thinks 'Claude has regressed' and that it 'cannot be trusted to perform complex engineering'

AMD's Senior Director of AI, Stella Laurenzo, has publicly documented a significant regression in Anthropic's Claude Code, a specialized AI coding assistant. In a detailed GitHub issue filed on April 2, 2026, Laurenzo presented an analysis of nearly 7,000 user sessions, revealing that the tool's performance has severely degraded since early March. Key metrics show Claude Code now reads three times less code before making edits, rewrites entire files twice as often, and has begun frequently abandoning complex tasks mid-way—a behavior that was previously rare. The decline appears linked to Anthropic's decision to completely redact the model's visible reasoning process, dropping its 'thinking aloud' from 100% to zero over just eight days in March.

This technical regression has had immediate business consequences. Due to the tool's newfound unreliability for complex engineering tasks, AMD's engineering team has already ceased using Claude Code and switched to a competing AI provider. Laurenzo urged Anthropic to restore the model's reasoning visibility and suggested the company introduce a premium service tier that guarantees deep, reliable reasoning. The performance issues coincide with a period of operational strain at Anthropic, which pushed out 14 rapid model releases alongside five service outages in March 2026, suggesting their quality assurance processes may be struggling to keep pace with aggressive development cycles. The public critique from a major tech company's AI lead highlights growing enterprise concerns over the stability and transparency of rapidly evolving AI models.