Claude Fable 5 is out. Here's what actually changed for people who ship software.
Anthropic released Claude Fable 5 today: the first Mythos-class model the public can actually use. If you've been following the Mythos Preview story since April, this is the "made safe for general use" version of the model that found a 27-year-old TCP bug in OpenBSD.
Here's the signal, separated from the launch-day noise.
The headline numbers
| Thing | Value |
|---|---|
| API model ID | claude-fable-5 |
| Input pricing | $10 / million tokens |
| Output pricing | $50 / million tokens |
| Subscription access | Free on Pro, Max, Team, and seat-based Enterprise until June 22; usage credits after |
| Safeguard fallback rate | < 5% of sessions |
That pricing is less than half of what Claude Mythos Preview cost partners. The API and consumption-based Enterprise plans get it immediately.
What's genuinely new
The capability claim is "longest task wins." Anthropic's own framing is that the longer and more complex the task, the larger Fable 5's lead over earlier models. The supporting anecdote is hard to ignore: Stripe reported a 50-million-line Ruby codebase migration completed in a day, work they'd scoped at two months for a team. It also posted the highest scores to date on Cognition's FrontierCode evaluation for production-quality code.
It's two models in one trench coat. Fable 5 and Mythos 5 are the identical underlying model. The difference is safeguards: Fable 5 ships with classifier gates on offensive cybersecurity, dual-use biology/chemistry, and distillation attempts. Mythos 5, with the cyber safeguards lifted, only goes to authorized security professionals through Project Glasswing.
The fallback is the interesting design choice. When a Fable 5 query trips a high-risk classifier, you don't get a refusal; you get an answer from Claude Opus 4.8 instead. Anthropic says this triggers in under 5% of sessions. For most software work you'll never see it; if you work in security tooling, you'll see it a lot, and that's the point.
Red-teaming held up, mostly. Over 1,000 hours of external testing found no universal jailbreak, though the UK AISI reportedly made initial progress on one. Treat the safeguards as a moving target, not a solved problem.
What it means for your sessions
A model that holds focus across millions of tokens and runs multi-hour autonomous tasks changes where the human leverage is. You spend less time fixing individual answers and more time steering between checkpoints, which is exactly the territory our deck of nudges was built for. We'll publish a fresh eval of the deck against Fable 5 once we've run it; until then we're not making claims. (That's a rule we learned the hard way.)
Keep Working is built by Urbano DX. We build AI software for clients: paid proof-of-concepts, production APIs, and AI workflows. We've been shipping with Claude models since well before they could migrate 50 million lines of Ruby. If your team wants to put Fable 5 to work, talk to us.