The 85% drop is a capability gain, not a safety retreat.
Anthropic says a rewritten classifier cut biology-related fallbacks by about 85% in its own testing. It expects total fallbacks to decline by roughly 67% on Claude.ai, 55% on Cowork, 17% on Claude Code and 7% on Claude Platform.
The boundary still matters: everyday health, education and some clinical-support requests should reach Fable 5 more often, but dual-use virology, toxicology and molecular-design work still routes to Opus 5. The numbers are vendor-reported, so production fallback rates—not the launch claim—are the next proof point.
Forward this to teams evaluating AI access controls in high-risk domains.
Source: Anthropic News · 7 August 2026
August 8, 2026 123 1
