SAN FRANCISCO, 8 AUG 2026 — If a blood-test result pasted into Claude has ever produced a noticeably vaguer answer than you expected, this is why: the question was being quietly handed to a less capable model. Anthropic said on 7 August it has cut that happening by about 85%.

The mechanism has a name most users never see. Fable 5, Anthropic's most capable model, runs behind a safety classifier for biology, and when the classifier decides a request is dual-use, the request falls back to Opus 5 instead. You still get an answer. It comes from a smaller model, and nothing on screen tells you the swap happened.

What changed this week is where the line sits, not whether there is one.

What Anthropic actually did

The company rewrote the classifier's constitution — what it calls "a collection of rules to help the model discern between safeguarded and allowed content." After gathering feedback from experts "both internal and external to Anthropic," it generated fresh training data from the new rules and retrained the classifier.

The headline result: "This update reduced biology-related fallbacks by about 85% across our product surfaces."

The examples Anthropic gives for what should now go through are ordinary and specific — "interpreting lab results, understanding symptoms, and learning about biology in an educational context." Healthcare professionals, it says, get more support on clinical tasks.

What that 85% is, and what it isn't

The number counts downgrades avoided. It does not measure answer quality, and it is not an accuracy figure.

The distinction matters. Before this week a question about your thyroid panel was not refused; it was routed to Opus 5. So the change is about which model answers, and the case for caring is that the two models are not equally good — which is Anthropic's own premise for keeping Fable 5 behind the classifier in the first place.

The figure is also Anthropic's, measured on Anthropic's product surfaces, and no independent party has replicated it. There is no published before-and-after on how often the classifier is wrong in the other direction, which is the number a cautious reader would want next.

Why retraining a classifier is the harder fix

It would have been cheaper to move a threshold. A classifier produces a score, and a company that wants fewer false alarms can simply require a higher score before it acts — one number, changed in an afternoon, fewer downgrades tomorrow.

Anthropic did not do that. It rewrote the rules the classifier is trained against, regenerated training data from those rules, and retrained the model. The difference matters because a threshold change trades the two kinds of error against each other in a straight line: fewer wrong blocks, more wrong passes. Retraining on better-specified rules aims to improve both kinds of error at once — to teach the classifier the difference between a worried patient and a virologist, rather than just making it less twitchy.

Whether it worked in that stronger sense is not something the published figure can tell you. An 85% drop in fallbacks is consistent with a sharper classifier and also consistent with a looser one. The measurement that would separate them — how often the new classifier lets through something the old one correctly caught — has not been published.

What is still behind the wall

Three domains remain gated: virology, toxicology and molecular design. A request the classifier reads as dual-use in any of those still falls back to Opus 5, and Anthropic is explicit that the update does not open up "professional biology research and drug development."

So the shape of the change is narrow and deliberate. Consumer and clinical questions get the better model; the research frontier does not. Anthropic says it is working on "a safe, scalable path for researchers to use our most capable models via trusted access pathways" — a vetted-access programme, in other words, which does not exist yet.

The three gated domains are not arbitrary. Virology covers making a pathogen more transmissible, toxicology making a compound more harmful, and molecular design is the step that turns either intention into a specific thing to synthesise. They are the points at which knowledge becomes a recipe, and they are the same three a researcher with an entirely legitimate project needs. That is the definition of dual-use. An access programme keeps being the answer because a better filter cannot solve it: the request looks identical whoever is typing it, so the only remaining variable is who they are.

The timing is doing a lot of work

One day earlier, Science published a study in which genome language models designed 16 working bacteriophages from scratch. We wrote about the numbers in that paper separately: the 16 working viruses came out of 285 that were physically built, and the successes clustered among designs that were near-copies of the template.

Put the two side by side and the week reads oddly. A chat assistant is being allowed to discuss symptoms more freely, while purpose-built genome models are demonstrating the capability that the assistant's virology gate exists to withhold. The gate on the chat interface is not what stands between anyone and a designed genome. Both things can be sensible at once, and they are also a reminder that consumer-facing safety controls and biosecurity controls are answering different questions.

What this changes for an ordinary user

Practically, three things.

You will hit fewer dead ends on everyday health questions, and the answers should be more substantive than the hedged versions the fallback tended to produce.

You still cannot tell which model answered you. There is no badge, no notice, and no log a user can check — so if an answer feels thinner than usual, "it fell back" remains a guess rather than something you can verify.

And none of this makes a general-purpose model a clinician. Anthropic's framing is that professionals get more support on clinical tasks, which is a claim about assisting people who already know what they are looking at. For everyone else, a model that explains what a lab reference range means is useful, but it is still not a diagnosis. Bring the result to someone who can act on it.

What to watch

Whether the trusted-access pathway ships, and who qualifies for it, because that is where the interesting policy decisions get made. Whether Anthropic publishes a false-positive rate to sit beside the 85%. And whether any competitor documents its own biology routing at all — Anthropic is unusual in describing the mechanism publicly, which is the reason this article can be written about its classifier and not anyone else's.