Since 7 August, Fable 5 answers health and study questions far more often. The change isn't in the model itself, but in the rules of the classifier that decides what counts as dual-use. Virology, toxicology and molecular design stay closed.
- By their own tests, biological escalations to the fallback model drop by about 85%. The other numbers - 67% for Claude.ai, 17% for Claude Code - are an expectation, not a measurement.
- The everyday stuff unblocks: reading research, understanding symptoms, learning biology. Dual-use stays blocked.
- The method is telling: they rewrite the classifier's rules, ask outside experts, build new data, retrain. They don't loosen the filter wholesale.
Since 7 August, Fable 5 answers health and study questions far more often, and biological refusals have fallen by about 85 per cent. The model itself is the same. They changed its guard.
What pleases me here, as a builder. Until now the safety conversation ran between two extremes - either you tighten until the tool is useless, or you loosen it and hope. Anthropic does the third thing: leave the wall in place and move the door - first rewrite the rule, then check it with people, and only then touch the data.
Ever tried to read up on research or figure out what a metric means, and hit a wall? That's now far rarer. Don't expect a difference in coding work: the expected drop there is 17%, meaning almost nothing. And watch the 85% number - it's for biological escalations, from their own tests, not from an independent measurer.