Anthropic's 186-page risk report runs on one number

Frontier AI labs now treat their own capability numbers the way central banks treat inflation prints: a single moving point is enough to redraw the policy line. Mowshowitz's reading of Anthropic's August Risk Report makes the pattern visible, and McAteer's X post pinpoints the number that does the work.

Anthropic's second voluntary risk disclosure is built around a curve. The company estimates that roughly 85% on CoBench v2, a suite of historical AI research tasks that Anthropic staff have already solved, is the threshold at which a model could replace its own researchers. The internal-only Model 2, kept off the market, sits 12.5 percentage points above Mythos 5 on that same suite. That gap is the document's actual headline. The 186 pages of taxonomy are the apparatus around it.

A capability number changes posture before it changes policy. Anthropic has already raised the Responsible Scaling Policy thresholds tied to AI-R&D automation and to novel biological and chemical risk, which means the next model jump will be read against a higher bar. The same numbers that prompt a lab to widen disclosure are the numbers that would, at scale, let one model do the science that builds its successor.

Mowshowitz frames the report as a "moderately positive update overall, if we presume they are not silently omitting the worst of it." That conditional is the read. A voluntary 186-page disclosure earns trust only at the level the disclosure itself names, and the 12.5-point jump names that level clearly.