Anthropic loosens Fable 5's biology filter: 85% fewer false blocks
Fewer health and lab prompts get downgraded to a weaker model — though the API gains least. Plus: an open MIT coding agent that beats the human baseline.
Copy markdown
Health and lab prompts stop getting downgraded
Anthropic retuned Fable 5's biology classifiers, cutting false-positive “fallbacks” to the weaker Opus 5 model by ~85% on biology queries. If your health, lab-result, or clinical-support prompts were coming back oddly degraded, that misfiring classifier was why — now far less trigger-happy.
The gain is uneven — and the API sees the least
Fallback reductions vary by surface: Claude.ai ~67%, Cowork ~55%, Claude Code ~17%, and Claude Platform (API) just ~7%. Build on the API and you'll notice the smallest shift. Genuine dual-use work — virology, toxicology, molecular design — stays blocked.
Elsewhere: an MIT agent that edges the human baseline
Prime Intellect open-sourced Prime Agent, a self-improving “recursive language model” coding harness that hit 95.5% on ARC-AGI-3 — just past the 95.4% human-expert bar. One-command install on Linux/macOS; runs on your Claude Pro/Max plan, API keys, or local models via Ollama or vLLM.