Anthropic will restrict a frontier model over theoretical cybersecurity risks, a problem addressed by Opus 4.7's built-in cyber safeguards. But it shipped a system prompt change to three production models with no apparent staged rollout, degraded Claude Code quality for weeks, and ran a pricing experiment that updated public docs globally while treating it as a small A/B test. Jonathan Neen's essay makes the case that these aren't separate categories. They're the same trust balance sheet.
The Mythos model stays restricted because Anthropic decided its cybersecurity capabilities crossed a line. That discipline cost real revenue. Meanwhile, Claude Code's April degradation came from a reasoning effort default change and a thinking-cache bug, compounded by a verbosity-reducing system prompt. The postmortem was transparent, as these things go. But Neen points out it grouped a deliberate product decision with actual bugs. Three weeks to revert a shipped default tells you everything about where judgment failed.
The pricing incident cuts cleaner. Amol Avasare, Anthropic's Head of Growth, called it a 2% test for new signups. The public pricing page changed for everyone. Simon Willison captured the core problem: when you treat developers as an A/B cohort instead of customers making procurement decisions, you don't lose subscribers. You lose trust. Fortune reported users describing the experience as feeling 'gaslit', mirroring the feedback described in reports that Claude is getting worse.
Once trust erodes, it doesn't come back. Hacker News sentiment shifted from glowing to openly cynical within days.
Anthropic's whole market position rests on being the lab that takes trust seriously. The Claude Constitution literally has a section on avoiding deception. Neen's real point is that Anthropic has drawn the safety boundary around model behavior, leaving product reliability and pricing as "just product." For developers building careers and toolchains around Claude Code, that distinction doesn't hold. Trust isn't compartmentalized.