June 10, 2026
Claude Fable 5 — The "Too Dangerous" AI Is Now Public
An AI beating Pokémon FireRed from raw screenshots alone. A working solar system built from scratch to predict eclipses. Apps that used to take 100 prompts, one-shotted. This is Claude Fable 5 — and two months ago Anthropic said this class of model was too dangerous to release.
What it is
In April, Anthropic described Mythos, a model so good at finding and exploiting software vulnerabilities they refused to ship it publicly. Fable 5 is Mythos — the same underlying model, with safeguards. Mythos 5, the unguarded twin, stays locked to cyber defenders. They call it a Mythos-class model: a new tier above Opus.
The numbers
- SWE-Bench Pro 80.3%, FrontierCode, ~90% on Hex
- Real builder results: a Stripe 50M-line migration, Cursor, vision-from-screenshots
- Safeguards that fall back to Opus 4.8 on risky tasks
- Pricing + free access before June 22 (details in the video)
Announcement → https://www.anthropic.com/news/claude-fable-5-mythos-5 · API model: claude-fable-5
Full first look + verdict in the video above. Subscribe for more on Claude and frontier models.