Claude Fable won't answer basic biology questions

Overview

Anthropic's newest public model, Claude Fable 5, refuses to answer basic biology questions on purpose. It is the first publicly available member of the Mythos model family, a class so capable at cybersecurity and biology that Anthropic considers the full version too dangerous to release. When a user asks about biology, chemistry, cybersecurity, or model distillation, Fable 5 hands the request off to the older Claude Opus 4.8 instead of answering itself.

Key Takeaways

Claude Fable 5, released June 9, 2026, is a safeguarded public version of Anthropic's Mythos-class model, which the company built but restricts to vetted users.
The model blocks itself from answering questions about biology, chemistry, cybersecurity, and distillation, even simple high-school-level biology, and routes them to Claude Opus 4.8.
The refusal is a deliberate safety design choice, not a knowledge gap; Anthropic decided overly broad blocking was worth releasing the model sooner.
A separate AI classifier system watches each request and triggers the handoff when it detects a sensitive topic, with users told whenever the switch happens.
Anthropic says the fallback is rare, with more than 95 percent of Fable sessions running entirely on the model itself.
Claude Mythos 5 is the same underlying model with fewer guardrails, available only to authorized cybersecurity professionals and biomedical researchers.

Stats & Key Facts

#More than 95 percent of Fable 5 sessions run entirely on the model with no fallback, per Anthropic's early data.
#Safety classifiers trigger a fallback in fewer than 5 percent of sessions on average, including harmless false positives.
#Anthropic ran an external bug bounty totaling over 1,000 hours of testing that produced no universal jailbreaks.
#Both Claude Fable 5 and Claude Mythos 5 launched on June 9, 2026.
#Anthropic set a mandatory 30-day traffic retention window for all Fable 5 users to defend against novel attacks.

Claude Fable won't answer basic biology questions

Why Claude Fable 5 Hands Off Biology Questions to Opus 4.8

The refusal is built in, not a limitation of what the model knows.

Ask Claude Fable 5 a routine biology question, the sort a high schooler handles, and it declines to answer directly. Instead a notice appears telling the user the request triggered a safety measure, and the response is produced by Claude Opus 4.8, Anthropic's previous flagship model.

The model is not missing the answer. Anthropic prevents Fable 5 from responding on a set of sensitive subjects by design. The company decided it was better to ship the model with broad, sometimes over-cautious blocking than to delay release while tuning the limits more precisely.

Mythos: The Model Family Anthropic Calls Too Dangerous to Release

Fable 5 belongs to a class of models Anthropic keeps mostly locked down.

›Mythos-class models are strong enough at finding and exploiting software flaws to make cyberattacks cheaper and easier to carry out.
›They show skill at agentic hacking, meaning they string together multiple stages of an attack such as reconnaissance, discovery, and lateral movement.
›Anthropic says the same models could give malicious actors meaningful help with high-risk biological research.
›Because of those risks, the company holds back the unrestricted version from general public access.

How the Safety Classifier and Fallback System Works

A second AI layer decides what Fable 5 is allowed to answer.

Anthropic runs separate classifier systems alongside Fable 5. These are dedicated AI tools whose job is to spot requests that touch cybersecurity, biology, chemistry, or distillation, which is an attempt to copy a model's capabilities into another system.

When a classifier flags a request, Fable 5 does not respond. The request is automatically routed to Claude Opus 4.8, and the user is told the switch happened. This keeps the most capable model away from the riskiest questions while still giving the user an answer from a safer model.

How Often the Handoff Actually Happens

Anthropic frames the blocking as an edge case rather than the norm.

›Early company data shows more than 95 percent of Fable 5 sessions run start to finish on the model itself.
›Fallbacks to Opus 4.8, including harmless requests caught by mistake, occur in fewer than 5 percent of sessions on average.
›Anthropic acknowledges the guardrails are broad enough to catch some innocent questions, which is the source of the basic-biology refusals readers notice.

Fable Versus Mythos: Same Model, Different Guardrails

The two releases share one engine and differ only in restrictions.

Claude Mythos 5 is the identical underlying model to Fable 5, but with certain safeguards removed. Access is limited to authorized users such as cybersecurity professionals and biomedical researchers rather than the general public.

Claude Fable 5 is the version anyone can use, wrapped in the classifier and fallback system. Anthropic positions Fable 5 as state of the art across nearly all of its capability tests, citing strong results on coding, finance, and cybersecurity benchmarks, while keeping the dangerous edges fenced off for everyday users.

What This Signals About Releasing Frontier AI Safely

The launch reflects a tradeoff between capability and caution.

The release landed days after Anthropic urged major AI labs to agree on coordinated limits for frontier development, warning that systems are advancing fast enough to approach self-improvement with little human input. Shipping a heavily fenced model fits that cautious posture.

For business readers, the practical takeaway is that the most capable models now arrive with built-in refusals on whole subject areas. Anthropic also stress-tested the guardrails with an external bug bounty exceeding 1,000 hours that found no universal jailbreak, and it requires a 30-day data retention window for all Fable 5 traffic to watch for new attacks.

Frequently Asked Questions

Why won't Claude Fable 5 answer basic biology questions?

Anthropic blocks the model from answering biology, chemistry, cybersecurity, and distillation questions by design. It is a safety choice, not a knowledge gap, and the blocking is broad enough to catch even simple high-school-level biology.

What happens when Fable 5 refuses a question?

The request is automatically routed to Claude Opus 4.8, Anthropic's previous flagship model, which answers instead. The user sees a notice explaining the safety measure triggered the switch.

What is the difference between Claude Fable 5 and Claude Mythos 5?

They are the same underlying model. Fable 5 is the public version wrapped in safety classifiers, while Mythos 5 has some safeguards removed and is limited to vetted users like cybersecurity professionals and biomedical researchers.

How often does Fable 5 hand off to Opus 4.8?

Anthropic says it is rare. More than 95 percent of Fable 5 sessions run entirely on the model itself, and fallbacks happen in fewer than 5 percent of sessions on average.

Why does Anthropic consider Mythos-class models dangerous?

These models are skilled enough at finding software vulnerabilities and at biological research that they could help malicious actors carry out cyberattacks or risky bio work. That is why the unrestricted version stays out of general public access.

Claude Fable 5 shows how Anthropic is shipping its most capable model to the public while deliberately walling off its riskiest abilities. The basic-biology refusals are the visible side effect of guardrails that route sensitive questions to the safer Opus 4.8.