AI Providers Pulling the Rug on Power Users

Anthropic sent notices to Claude Code subscribers saying third-party harnesses like OpenClaw are no longer allowed under the subscription tier, offering a one-time credit as consolation. The thread lit up with founders asking why a usage-capped subscription can constitute 'outsized strain' and whether this is just a money grab dressed up as policy. Separately, a dedicated site called Livenerf is now tracking whether Opus 5.5 has been degraded, with commenters noting Sonnet 5 felt like a step backward and worrying the same will happen to the current flagship.

The pattern here: AI providers are in a bind. They priced subscriptions to attract power users, those power users are actually using them, and the compute math doesn't work. The response is quiet capability throttling, policy changes that penalize the heaviest users, and opaque 'strain' language that gives no recourse. Builders who built workflows on these subscriptions are now finding their implicit contract wasn't what they thought.

The counterpoint raised in the threads is fair: providers do need to manage capacity, and API keys with usage caps would be a cleaner solution than banning tools. But no one is offering that alternative. The practical lesson is that any workflow that depends on a flat-rate subscription to an AI provider is sitting on a fragile foundation.