A Claude reliability thread argues that recent “Claude got dumber” complaints may be real, but not necessarily caused by the base model degrading. The more enterprise-relevant explanation is scaffolding drift: cache TTLs, adaptive thinking settings, and effort-routing changes altering outputs in production.
The cache angle matters because it links quality complaints to cost and latency behavior, not just model intelligence. For CTOs, this reinforces the need to monitor prompt cache hit rates, TTL changes, and vendor-side behavior shifts as first-class reliability metrics.
The “adaptive thinking” and “effort flip” claims point to a bigger operational risk: the same prompt can receive different reasoning budgets depending on hidden routing or configuration. Enterprise AI teams should test critical workflows against variance over time, not only benchmark them once before launch.
The thread connects directly to Niels’ earlier Anthropic bill spike investigation from 2026-04-13: unexpected spend and perceived quality drops can share the same root cause. AI leaders should treat model wrappers, orchestration, caching, and vendor defaults as part of the production system, not as neutral plumbing.
Today’s theme: AI reliability failures are increasingly less about the model in isolation and more about the invisible execution layer around it.