The Best Model Won't Do Your Security Review

I asked the most capable model Anthropic makes to review my code for security holes. It politely declined - and handed the job to a lesser one. The lesser one walked my code and found a chain of flaws I'd have been proud to catch myself.

Here's the setup. Fable 5 is Anthropic's flagship - the sharpest, most expensive model they offer. Opus 4.8 is the everyday workhorse, the one most of my work actually runs on. So I did the obvious thing and pointed the sharpest tool at the hardest problem: a real security review of a real codebase. What I got back wasn't a refusal, exactly. It was a handoff. Fable stepped back, Opus stepped in, and the review ran on the cheaper model without me asking.

This has a name. Anthropic calls it a safeguard reroute. Fable runs classifiers over your request, and when it lands in a sensitive, dual-use domain - cybersecurity, biology and chemistry, or distillation (extracting a model's abilities to train a rival) - it quietly routes the request to Opus 4.8 instead. You don't even pay Fable's premium for it. The most capable model doesn't refuse the dangerous-adjacent work. It delegates it, downward.

Sit with how backwards that feels. The sharper the tool, the more they lock the blade - because a great security review is one keystroke from a great attack plan, and the same capability that finds your flaw can write the exploit. So Anthropic handicaps its best model exactly where raw power is most dual-use, and lets the workhorse carry the sensitive load. I'm not the only one who noticed the flagship is the clamped one; another builder put it plainly this week - Opus 4.8 let them go a little farther than Fable did.

And the workhorse delivered. Opus didn't just cope with the review - it found the real thing, a chain where one soft check let the next one through, the kind of finding I'd have called Fable-level if you hadn't told me which model produced it. Which is the question I actually came away with: if the step-down model does step-up work, is Opus quietly becoming Fable?

As far as I can tell, no. Opus is simply strong: Anthropic says more than 95% of Fable sessions never trigger a fallback at all, and Fable still wins overall. The gap closes the ordinary way, through new models - Opus 5 is already leaking - not through a silent swap. And Fable itself isn't headed for pay-per-use exile forever; Anthropic says the metering is temporary, back to normal plans once capacity allows.

The flagship refused the job. The workhorse did it - and found the chain.