Anthropic has adjusted the biology safeguards around Claude Fable 5 after finding that the system was sending too many ordinary questions to a more restrictive fallback model. The company says its revised classifier reduced biology-related fallbacks by about 85 per cent in testing, while continuing to intercept prompts that could support harmful or dual-use work.
The update is less about expanding Fable's scientific frontier than making its existing capabilities easier to use. Questions about common health concerns, symptoms, lab results and biology lessons should now be less likely to trigger an unexpected hand-off. Anthropic also expects healthcare professionals to receive more useful clinical support from the model, although the announcement does not turn Claude into a medical authority or replace professional judgement.
A narrower boundary for everyday biology
Claude's fallback system acts as a gatekeeper. When a prompt appears to involve sensitive biological knowledge, it can move the conversation from Fable 5 to Opus 5. That is intended to preserve useful assistance while applying stricter treatment to requests that might help someone design or carry out biological harm.
The earlier version of that gate was catching too much. Benign prompts could be swept into the same path as genuinely sensitive ones, interrupting conversations and making the model feel inconsistent. Anthropic's reported reduction therefore measures fewer unnecessary interventions, not a blanket relaxation of biological safety controls.
The practical effect varies across Anthropic's products. The company estimates that overall fallbacks will decline by roughly 67 per cent on Claude.ai and 55 per cent in Cowork. The expected reductions are smaller in Claude Code, at about 17 per cent, and on Claude Platform, at about 7 per cent. Those figures reflect different patterns of use and do not amount to a promise that every biology question will stay with Fable.
Some scientific work still belongs elsewhere
Anthropic is explicit about the line that remains. Prompts involving dual-use fields such as virology, toxicology and molecular design can still be routed to Opus 5. These are areas where legitimate research can overlap with knowledge that could increase biological risk, so the company is keeping a more cautious path for them.
Fable 5 is also not positioned for professional biology research or drug development. That limitation matters because a smoother response to everyday health and education questions could otherwise be mistaken for a broader scientific endorsement. The update improves access to lower-risk help; it does not certify the model for advanced laboratory, discovery or clinical decision-making.
For organisations using Claude in health-adjacent settings, the distinction is important. Fewer false positives may reduce friction for staff and customers, but existing review, privacy and safety processes still apply. Anthropic has not announced new compliance coverage, diagnostic approval or a change to the responsibility users carry when acting on model output.
Anthropic retrained the gatekeeper
The company says it rewrote the classifier's constitution, produced updated training data and retrained the system before evaluating it. In this context, the constitution is the set of principles used to teach the classifier how to separate harmless biological discussion from requests that require stronger safeguards.
That approach targets the decision layer rather than simply adding exceptions for a list of common questions. A classifier trained on clearer boundaries should be better able to recognise context: a student asking for an explanation of a cell process is different from a request for operational help with a hazardous biological agent, even when some vocabulary overlaps.
Anthropic says the revised system continues to block harmful requests and preserve protections around dual-use research. The company has not published enough detail in this announcement for outsiders to reproduce the 85 per cent result, so the figure should be read as Anthropic's own evaluation rather than an independent benchmark.
A smoother experience without removing the hard limit
The most visible outcome should be continuity. Users asking ordinary biology or health questions are less likely to see the model change underneath them, while high-risk topics retain a separate route. That is a useful refinement because safety mechanisms work best when they are precise enough that people do not encounter them constantly in harmless situations.
It also shows the trade-off in running model-level safeguards across several products. One classifier has to serve consumer chat, collaborative work, coding and API use, each with different prompt patterns. Anthropic's product-by-product estimates acknowledge that the same retraining will not have an identical effect everywhere.
The uneven reductions also provide useful context for the headline result. An 85 per cent decline in biology-related fallbacks in testing does not mean every Anthropic surface will show an 85 per cent change in total fallbacks. Product mix, the kinds of prompts people submit and the safeguards already applied in each environment all affect what users ultimately encounter.
The update leaves Fable's central constraint intact: it can support ordinary learning and lower-risk questions, but it is not the model Anthropic wants professionals to rely on for advanced biological research. The change makes that boundary less intrusive, not less consequential.