The AI you got today isn't the real one.
On June 10, 2026, Anthropic launched Claude Fable 5 — the most capable model they've ever released to the public.
But here's what's buried in the announcement:
The actual full-power version — Claude Mythos 5 — runs on the same brain. Same model weights. Same architecture.
You just don't get it.
Why?
Because Anthropic found it could autonomously execute multi-stage cyberattacks. Perform gene therapy virus design that outperformed dedicated protein AI models. Generate insights that could give bioweapons researchers a meaningful leg up.
So they locked Mythos 5 behind Project Glasswing — a US government program for vetted cyber defenders — with a biology trusted access program for researchers also planned.
What you DO get — Fable 5 — is still historic:
→ Stripe ran a 50-million-line Ruby migration in 1 day. Their team estimated 2+ months by hand. → It beat Pokémon FireRed using vision alone. No maps. No tools. Just raw screenshots. → It designed real drug candidates across 9 of 14 protein targets now under active investigation. → It costs less than half of the previous frontier Claude model.
There are 3 classifiers running on every Fable 5 query: cybersecurity, biology/chemistry, and model distillation. Trip any of them and you're handed off to a different model — and you'll be notified when this happens.
95%+ of sessions never trigger them.
But the 5% that do? That's where the ceiling is.
One more thing: free access on subscription plans ends June 22. After that — usage credits only.
The most powerful AI ever given to the public comes with a leash. And the leash has a very deliberate shape.
