Anthropic’s new Claude Fable 5 blocks basic biology questions with safety filters

Anthropic is marketing Claude Fable 5 as its most powerful generally available model, but the new system is already exposing a sharp gap between promise and practice. Biology questions that a high school student could handle are being blocked by safety filters and sent to Claude Opus 4.8 instead, a reminder that frontier AI can still stumble on routine work even as it is sold as ready for serious use.

The company said Fable 5 launched June 9 and is the same underlying model as Claude Mythos 5, but with added safeguards for general release. Anthropic says those safeguards cover biology and cybersecurity, and that flagged queries are automatically routed to Opus 4.8. The company also says the filters trigger, on average, in less than 5% of sessions. For schools, researchers, and employers trying to build AI into lesson plans, lab workflows, and internal knowledge systems, that means a model advertised for biology may still refuse ordinary questions or quietly shift them to a different engine.

Pricing and access rules reinforce the split between public-facing capability and restricted scientific power. Fable 5 and Mythos 5 are both priced at $10 per million input tokens and $50 per million output tokens, while U.S.-only inference is available for Fable 5 at 1.1 times that price. Mythos 5 is initially limited to a small set of vetted partners, first through Project Glasswing, with a broader trusted-access program planned later. Using Mythos 5 requires accepting a 30-day data retention policy for safety monitoring.

Anthropic has spent the past year pushing harder into life sciences. It added Claude for Life Sciences in October 2025, expanded healthcare and life-sciences tooling in January 2026, and said Opus 4.5 showed gains in figure interpretation, computational biology, and protein-understanding benchmarks. Its life-sciences product page now says Opus 4.8 brings broader, more reliable reasoning to chemistry, structural biology, bioinformatics pipelines, and lab protocol design.

That positioning sits alongside a more aggressive security push. Project Glasswing launched in April 2026 with partners including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic says about 50 partners using Claude Mythos Preview found more than 10,000 high- or critical-severity vulnerabilities, and on June 2 it expanded the program to about 150 additional organizations in more than 15 countries. The company is making a clear choice: keep its strongest biology and cybersecurity capabilities under tighter controls, even as it presents Fable 5 as the model most people can use.