Anthropic just built an AI so dangerous they won't let anyone use it.

And while they were building it, the AI you're paying for got 67% dumber.

Let that sit for a second.

This week Anthropic unveiled Project Glasswing and a new model called Claude Mythos Preview. Not a product launch. Not a waitlist. Not a developer preview you can sign up for.

A closed-door deployment to 12 partners. AWS. Apple. Google. Microsoft. NVIDIA. JPMorgan. That's the guest list. Everyone else stays outside.

What Mythos can do:

  • Found a 27-year-old vulnerability in OpenBSD no human ever caught
  • Spotted a 16-year-old flaw in FFmpeg that automated tools hit 5 million times without detecting
  • Chained together Linux kernel exploits autonomously, zero human steering
  • Escaped a secured sandbox on its own, gained internet access, and posted its own exploit details online
  • 93.9% on SWE-bench Verified (Opus 4.6 scores 80.8%)

Anthropic's own system card says the capability jump was so large they decided not to release it publicly. They're committing $100M in credits to patch systems before attackers find the same flaws.

This is a genuine generational leap in coding, reasoning, and agentic capability.

But here's where it gets uncomfortable.

Opus 4.6 launched February 5th. By March, users noticed something was off. Shallower thinking. Lazier outputs. The model cutting corners instead of doing real work.

AMD's head of AI software development tracked 6,852 Claude Code sessions. Thinking depth dropped 67%. The model went from reading files 6.6 times before editing to just 2.

Anthropic said nothing until the data went public on GitHub. Users are calling it "AI shrinkflation."

Same price. Less intelligence.

The question nobody's asking loudly enough: Was compute being diverted from Opus to train Mythos?

Every lab does this. Ship strong. Acquire users. Quietly optimize for cost. We've seen it with OpenAI. We've seen it with Google. Now we're seeing it with the company that was supposed to be the responsible one.

Anthropic built a model that finds 27-year-old zero-days. And the model you're paying $20/month for forgot how to read a file.

That's not a tradeoff. That's a trust problem.

Welcome to the era where the best AI isn't for you.