Production teams are reporting that “Claude got dumber,” but the strongest signal points to orchestration issues rather than a pure model regression. The Reddit analysis cites cache TTL behavior, adaptive thinking changes, and effort-level flips as likely causes, which matters because enterprise AI failures may originate in configuration drift, not vendor model quality.
AI reliability now depends as much on scaffolding as on the base model. For CTOs and AI leads, this means observability must cover prompt routing, caching, tool calls, reasoning settings, and fallback logic, not just output quality scores.
The pattern echoes earlier enterprise cost surprises, including Niels’ April investigation into an Anthropic bill spike. The lesson is that small runtime policy changes can hit both quality and spend, so AI governance needs release controls, regression tests, and cost anomaly detection around every orchestration layer.
Today’s theme: enterprise AI risk is shifting from model selection to operational control of the systems wrapped around the model.