Anthropic Cut Fable 5's Biology Refusals 85%. The API Barely Moved.

TL;DR

Anthropic shipped a fix today for the single most mocked thing about Claude Fable 5: the biology classifier that spent two months refusing to discuss mitochondria. Per the announcement, the company rewrote the classifier's constitution, retrained it on new data, and cut biology-related fallbacks by roughly 85%. Total fallbacks drop about 67% on Claude.ai and 55% on Cowork. On Claude Code the number is 17%, and on the Claude Platform, the API you actually build against, it is 7%. That gap is the story.

What actually changed

Fable 5 does not refuse in the usual sense. It runs input classifiers on every request, and when one fires, the request is silently handed to a weaker model that answers instead. Anthropic calls this a fallback. Today's change did not adjust a sensitivity threshold. It replaced the document the classifier reasons from.

Anthropic describes rewriting the classifier's constitution, which it defines as a collection of rules that help the model discern between safeguarded and allowed content, "taking care to carve out benign uses in detail." The company says it solicited feedback from outside experts, built new training data from the rewritten rules, and retrained.

A constitution here is the written rulebook the guard reads before deciding. The old one said, roughly, turn away anyone carrying something sharp, so the door staff dutifully turned away a chef holding a butter knife. The new one enumerates which knives are fine. The model that would not explain mitochondria in June will explain them in August.

The relief is not evenly distributed

Anthropic published one biology number and four surface numbers, and the arithmetic between them is more informative than either alone. Biology fallbacks fell about 85%. If that cut landed evenly across surfaces, then biology has to account for roughly four fifths of every fallback on Claude.ai (0.67 divided by 0.85) and roughly one twelfth of every fallback on the Claude Platform (0.07 divided by 0.85). Anthropic did not publish that breakdown, so treat it as arithmetic rather than a disclosure.

The implication is still hard to dodge. Consumer chat traffic that trips the classifier is mostly people asking about symptoms, lab results, and how cells work. API traffic that trips the classifier is mostly something else, and that something else did not get touched today.

For the Claude Platform that works out to a seven percent cut, the kind of improvement you notice only if you were already charting it.

What still falls back

Four categories still route requests away from Fable 5, per Anthropic's help center documentation, updated today:

• Offensive cybersecurity, meaning exploits, malware, and attack tooling. These fall back to Claude Opus 4.8.
• Dual-use biology and chemistry. Anthropic names virology, toxicology, and molecular design specifically, and says Fable "isn't yet usable for professional biology research and drug development." These now fall back to Opus 5.
• Distillation attacks on Fable 5, including attempts to extract its summarized thinking.
• Frontier LLM development, such as distributed training infrastructure work.

If you are writing security tooling, doing red-team work, or building training infrastructure, none of that moved. The cyber classifiers have their own separate lineage, described in Anthropic's July 2 cyber safeguards and jailbreak severity framework post, and they were not part of today's retrain.

How this shows up in your code

Worth knowing regardless of today's numbers, because the mechanism is not going anywhere. When the classifier fires and you have no fallback configured, the API returns stop_reason: "refusal" with a stop_details object carrying a category and an explanation. Input tokens on a direct block are not billed.

If you opt into server-side fallback, the response comes back under the fallback model's id with an extra content block of type fallback recording the from and to models, and a usage.iterations array with one entry per model that ran. Anthropic's documentation calls usage.iterations the source of truth for a definitive per-turn check, which is the polite way of saying do not trust the streaming event alone.

The billing detail is the one people miss: on a Fable to Opus fallback, the input tokens for the second call are billed as a cache read at roughly 10% of base price rather than a fresh cache write. That happens automatically with server-side fallback, and via a credit token with a five minute validity window if you are doing it client side. Details are in the fallback and billing guide.

Two months from apology to fix

Fable 5 launched June 9 as the first publicly available Mythos-class model, and Anthropic's own launch post claimed that "more than 95% of Fable sessions involve no fallback at all." Within a day that number was colliding with reality in public. The Register catalogued users hitting fallbacks on the word "cancer," on a security architect's resume, and on ordinary machine learning work. Anthropic said within 24 hours that it had "made the wrong tradeoff" and would make refusals visible rather than silent.

It took roughly two months to go from that apology to a retrained classifier. That is not slow for a system where the failure mode on the other side is uplift for someone building a pathogen, and Anthropic is explicit that it deliberately erred conservative. It is also two months during which the most capable public Claude was, for a large slice of legitimate scientific work, quietly not the model answering.

The company says it remains "fully committed to developing a safe, scalable path for researchers to use our most capable models via trusted access pathways," and the help center adds that it will consider opening allocations for dual-use cyberdefense and biology research, with program details to come. No dates.

What to do with this

If you have a health, education, or general science product on Claude and you benchmarked it against Fable 5 before today, rerun your evals. The model your users hit changed underneath you, and it changed in the direction of Fable answering more often, which means better answers and a different cost profile than a fallback to Opus.

If your workload is security tooling, drug discovery, or training infrastructure, nothing changed. Keep your fallback path wired and keep instrumenting usage.iterations, because you are still going to hit it.

Key Takeaways

• Anthropic retrained Fable 5's biology classifier against a rewritten constitution, cutting biology-related fallbacks by about 85%.
• Total fallback reductions are lopsided: roughly 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on the Claude Platform.
• The arithmetic implies biology was most of what tripped the classifier in consumer chat and only a small slice of what trips it on the API.
• Dual-use biology (virology, toxicology, molecular design) still falls back to Opus 5. Offensive cybersecurity still falls back to Opus 4.8. Distillation and frontier LLM development still fall back too.
• Fallback input tokens bill as a cache read at about 10% of base price, and usage.iterations is the reliable way to detect that it happened.
• Trusted access pathways for professional biology research are promised but undated.

Sources: Anthropic, Improving Fable 5's biology safeguards, Claude Help Center, Why Claude switched models in your conversation with Fable 5, Anthropic, Claude Fable 5 and Claude Mythos 5, Anthropic, Fable safeguards and jailbreak severity framework, Claude Cookbook, Classifier fallback and billing for Claude Fable 5, The Register