What Is Actually New

The fix sits in the classifier, the layer that reads each biology prompt and decides whether Fable 5 answers directly or falls back to the safer Opus 5. Anthropic rewrote that classifier's "constitution," the plain-English rulebook it follows, after taking feedback from internal reviewers and outside experts.

The payoff is fewer false alarms, though how many depends a lot on where you use it. Biology fallbacks are down roughly 85% overall, but broken out by product that's about 67% fewer on Claude.ai, 55% in Cowork, 17% in Claude Code, and only 7% on the Claude Platform. In everyday terms, Fable 5 will now walk you through a lab report, talk through symptoms, help with schoolwork, and back up clinicians on routine questions instead of refusing.

What Still Gets Blocked

The genuinely risky requests stay fenced off. Anything touching virology, toxicology, molecular design, or professional drug development still routes to Opus 5, which Anthropic treats as its safer default for dual-use science.

The classifier now sorts prompts into tiers. Clearly harmful content is blocked, and so is legitimate dual-use research. There's also a buffer of low-risk material that gets blocked anyway, on purpose, as a margin of safety. Whatever it reads as plainly harmless goes through. So the net got smaller, not removed.

Who This Actually Helps

The loudest complaints came from people with the most mundane needs: students, patients trying to understand their own results, and healthcare workers who kept hitting a wall in the middle of a task. Those are the exact queries this update aims at. If you were the one getting refused for asking what your cholesterol number meant, this is for you.

One Thing to Watch

That deliberate buffer zone is the part I'd keep an eye on. Anthropic is openly saying it blocks some harmless requests on purpose, trading a little usefulness for a safety cushion it can't calibrate perfectly. Where exactly that line falls is a judgment call, and judgment calls tend to move once real usage data comes in. Fable 5 also shipped next to Mythos 5, the same model with more guardrails removed for vetted security work, so the company clearly has room to run any given version tighter or looser. The open question is whether the next tweak loosens things further, or whether Anthropic decides 85% already went too far.