Anthropic just released the model they said was too dangerous

Anthropic just released the model they said was too dangerous to make public.

For once, the dramatic backstory is true. Whether it matters to your business is a different question.

In April they built an AI so good at finding security flaws they refused to release it. They handed it to security teams at companies like Microsoft and Google instead, so the holes could be fixed before anyone exploited them.

That model is now public, with guardrails. It's called Claude Fable 5. Ask it about cyber attacks or bio-weapons and the request is blocked or rerouted to an older model. For everything else, it's the most capable AI anyone can use.

I run my business on Claude every day, and this changes less than the coverage suggests. The everyday models are still faster and cheaper for emails and quotes. Paying double to reword an email gets you the same email.

Fable 5 is built for a different class of work. It can hold one problem for hours, even days, without supervision. Think of a full tender response, a year of accounts analysed for patterns, or a working piece of software for your business. Jobs you'd block out a full week for.

On the API it costs double Opus 4.8, their previous best, and on paid Claude plans it burns usage at twice the normal rate. Anthropic built this for big jobs and priced it that way.

Most of us have never given AI a job that takes longer than ten minutes. The brief became the limit this week, because the model can handle far more than we think to ask.

Write down the one piece of work you keep putting off because it would take a week. That's the first job for Fable 5.