After months of caution over AI security risks, Anthropic has broken the silence with Claude Fable 5.1, a sharper, cheaper, and less restrictive model. OpenAI’s Astra, the first to cross its “critical cybersecurity threshold”, is due this week with tighter access to advanced cyber features. The frontier-model race is back on.
Fable 5.1 delivers big jumps in coding, long research tasks, and knowledge work, with Anthropic claiming up to 25% lower cost on typical jobs. Its safety filters now intervene far less: 60% fewer false positives on cybersecurity work and 85% fewer on basic biology and medical questions. This fixes a major user complaint from Fable 5.
Alongside it, Anthropic launched Mythos 5.1—the same base model but with lighter guardrails for vetted US researchers in cybersecurity and life sciences. This two-track approach (broad Fable vs restricted Mythos) mirrors OpenAI’s plan to gate Astra’s most powerful cyber abilities. Same model, different safety tiers for different users is the new norm.
The road ahead points to calibrated access, not blanket bans: companies will ship powerful models but restrict risky features to verified partners. Not all stakeholders are fully aligned—governments, researchers, and firms are still negotiating how much capability is safe to release. The “safety over shipping” pause appears over, replaced by guarded, tiered rollouts.
AI IS NO LONGER HELD BACK—IT IS BEING GATED, GRADED, AND SENT OUT WITH GUARDS.
Sanjay Sahay
Have a nice evening.