AMD's AI Director Just Proved Claude Code Is Getting Dumber — 6,852 Sessions of Data Don't Lie

TamilTech AI summary

AMD’s AI Director Stella Laurenzo dug into 6,852 real Claude Code sessions, hundreds of thousands of tool calls, and thinking blocks from January through early April 2026 and found the tool got measurably worse, not better. After Claude Code v2.1.69 rolled out with thinking-content redaction, stop-hook violations spiked, careful file reading collapsed from about 6.6 reads per edit down to roughly 2, and the model started rewriting whole files instead of making tight targeted patches, which burns more tokens, risks more bugs, and messes up reviews and Git history. Prompt-caching bugs also quietly inflated some costs by 10–20x, which hurts enterprise and tight-margin teams hard, and major outages around April 6–7 only made the reliability worries louder. This matters because so many developers—and big Indian IT shops and startups—lean on AI coding assistants daily, so silent quality drops and surprise bills can hit deliverables and budgets. What you should know: don’t bet on one tool alone, watch your billing, always human-review AI changes (especially full-file rewrites), track whether the assistant is actually helping, pin versions before auto-updating, and treat the whole thing as trust-but-verify rather than blind reliance.

• What did AMD find wrong with Claude Code?
• What caused Claude Code to degrade?
• Should developers stop using Claude Code?
• How does this affect Indian IT companies?

AI-assisted summary, checked by the TamilTech editorial team.