I've been building with Claude subscriptions for months. This week felt like a wake-up call.

Since March 23, Claude Max users (including me) have been hitting session limits in 1 to 2 hours instead of the full 5-hour window. Anthropic called it their top priority. Still not fully fixed.

Then on April 4, subscriptions stopped covering third-party agent tools like OpenClaw. Users facing up to 50x cost increases. The explanation: flat-rate subscriptions weren't designed for 24/7 autonomous usage. That's fair, but the timing (locking out open source tooling while building your own closed harness) makes the motivation readable in more than one way.

Meanwhile, Anthropic announced Claude Mythos Preview. Enterprise-only, through Project Glasswing. It found a 27-year-old OpenBSD vulnerability and a 17-year-old FreeBSD remote code execution flaw, autonomously. Opus 4.6 comparison on Firefox exploits: 2 vs 181. Over 99% of findings still unpatched. Keeping it locked is the right call. The fact that this decision had to be made at all is the more interesting signal.

There's a lot happening in AI right now. Second post in my AI Opinions series on Digital Thoughts. Covers Mythos, Project Glasswing, Managed Agents, Meta Muse Spark, and what I'm testing this week.