Claude 5.1 System Card: Paper Risks vs. Real Ability to Break the Vase

200 Pages of Audit for the Top Model

Anthropic published the System Card for its new generation of models, Claude Fable 5.1 and Mythos 5.1. The document is over 200 pages long and maps out safety evaluations and model capabilities in detail. According to Zvi Mowshowitz, Fable 5.1 was, at the time of release, by a healthy margin the most capable publicly available AI model in the world. The report contains traditional risk breakdowns from CBRN to autonomous replication and shows that even with this advanced generation, the models do not cross the critical red lines set by ASL-3 (AI Safety Level 3).

When Paper Bureaucracy Misses Reality

For enterprise clients and security teams, this document highlights a growing problem in the form of compliance theater. The quantity of text does not translate to the practical utility of the findings. The report spends dozens of pages analyzing hypothetical and extreme scenarios, but it lacks a sharp delineation of everyday failure rates. Anthropic is building an airtight defense for regulators, but for practical application deployment, 200 pages serve more as an alibi than a guide on what exactly the model will screw up in a normal corporate routine.

The Model Is Safe Because It Cannot Plan

Celebrating successful safety tests comes with a catch. The models often fail to exhibit dangerous behavior simply because they fail at consistent, long-term planning. Zvi Mowshowitz aptly notes that the models don't threaten the world not because of perfect alignment, but because of their inability to execute a complex sequence of actions without getting lost along the way. This is a buffer for safety, but it is simultaneously a hard limit for truly agentic, autonomous work.

What Real Autonomy Will Reveal

The tipping point will not occur when the System Card grows to 500 pages. The critical shift will be the transition to models that maintain context and execution capability across hundreds of steps without human oversight. Once a model appears that can provably manage a long process stably without crashing, we will see whether current safety protocols can keep it within bounds, or whether past safety was merely a byproduct of technical imperfection.

Lilith's verdict

Bureaucracy won. While Anthropic prints 200-page insurance policies for regulators, product managers still have no idea at which step the model will break their actual corporate workflow.