Claude Fable 5 eases biology guardrails, cutting false refusals 85%
Everyday health and lab-result questions now go through; virology and toxicology stay blocked. Plus: OpenJDK bars all AI-written code.
Copy markdown
85% fewer biology fallbacks
Anthropic retuned Claude Fable 5's biology safety classifier and cut biology-related fallbacks by about 85% across its product surfaces. Everyday asks it used to refuse — interpreting your lab results, understanding symptoms, learning biology for a class — now get a real answer instead of a canned deflection.
Where you'll actually feel it
The drop is uneven by surface: Claude.ai sees 67% fewer fallbacks and Cowork 55%, but Claude Code gets only 17% and the Claude Platform API just 7%. If you build on the API, expect the old guardrails to keep biting far more than the headline number suggests.
Virology and toxicology stay walled off
This is a precision retune, not an open door: genuinely dual-use requests — virology, toxicology, molecular design — still bounce to the stricter Opus 5 classifier. Professional pathogen research and drug-design work remain off-limits.
OpenJDK still won't take AI-written code
Trending again on HN today: Oracle's OpenJDK interim policy bans any contribution "generated, in part or in full" by LLMs — even hand-editing 10 of 100 AI lines is barred, while Oracle's own GraalVM welcomes AI-assisted patches. Check a project's stance before you try to upstream Claude Code or Copilot output.