AMD abandons Claude Code over reasoning collapse

Anthropic's attempt to optimize compute costs by silently nerfing Claude Code's reasoning depth is backfiring spectacularly with power users.

• –An audit of nearly 7,000 sessions shows the model now reads code 3x less before editing, leading to sloppy, context-blind rewrites.
• –The regression directly correlates with Anthropic's late-February update that redacted visible "thinking" blocks, heavily implying a shift toward cheaper, shallower inference.
• –Reduced reasoning paradoxically increases API costs through "thrashing," as the model gets stuck in loops and requires constant human intervention to fix broken logic.
• –The incident highlights the growing tension between AI providers managing massive infrastructure costs and developers demanding guaranteed reasoning depth for complex agentic tasks.