AI News: Anthropic Launches Most Capable AI Yet, But You Can't Always Use It

The reason? Fable 5 is too good at hacking.

According to Anthropic's announcement, Fable 5 beats every model they've ever released on nearly all capability benchmarks. Software engineering, scientific research, vision tasks, knowledge work. The longer and more complex the task, the bigger the gap between Fable 5 and everything else.

During early testing, Stripe reported that Fable 5 compressed what would have taken a full team two months into a single day. The model performed a codebase-wide migration across 50 million lines of Ruby code autonomously. GitHub called it "a real step forward" for developers. Cursor said it "opened up a class of long-horizon problems that were out of reach for earlier models."

But Anthropic also admits something rarely acknowledged in AI announcements: this level of capability is dangerous. Without safeguards, Fable 5's cybersecurity skills could make cyberattacks "substantially easier and cheaper to commit."

Their solution is a system of AI classifiers that detect potentially risky requests and automatically downgrade the response to Opus 4.8. Anthropic claims this happens in less than 5% of sessions. They also admit the safeguards are "tuned conservatively" and will sometimes catch harmless requests.

For UK small businesses, this matters in two ways.

First, you're getting a more capable model at a lower price. Fable 5 costs $10 per million input tokens and $50 per million output tokens, less than half the price of Anthropic's previous top-tier model. If you're using Claude for software development, document analysis, or complex knowledge work, you're getting better results for less money.

Second, you're seeing the beginning of a new pattern in AI releases. The most capable models won't always be fully available to everyone. Anthropic is also launching Claude Mythos 5, the same underlying model as Fable 5 but with safeguards lifted. Initially, only cyberdefenders and infrastructure providers working with the US government will have access.

This isn't necessarily a bad thing. The alternative is either not releasing capable models at all, or releasing them with no restrictions and hoping for the best. But it does mean that as AI gets more powerful, access becomes more controlled.

The practical takeaway for small businesses: Claude Fable 5 works exceptionally well for legitimate business tasks. Software development, financial analysis, document processing, visual tasks. If you're already using Claude, the upgrade will be noticeable. If you're not, this is a good time to test whether AI automation can cut time and cost from your operations.

Just don't expect it to help you pen-test your own systems. For that, you'll get the previous model.

Want to explore how AI like Claude Fable 5 could automate parts of your business? Contact Aucta AI to discuss your specific use case.