Anthropic Says AI-Driven R&D Acceleration Has Met Its Safety-Policy Threshold

What happened

The determination appears in the second amended prospectus (S-1/A) filed on Nov 10. The previous version, dated Nov 2, said the company "may imminently meet" the threshold; the new version says it has been met. Under version 3.4 of the Responsible Scaling Policy, in force since July 8, the threshold is triggered when AI's contribution to its own development approximately doubles the rate at which models improve, requiring stronger safeguards; the company says it is evaluating the implications and will release further reports.

The filing says the determination rests on internal estimates for future models and the overall rate of progress, rather than on observed performance of current models; the system card for Claude Fable 5.2, dated Nov 9, says internal use of that model does not accelerate the company's aggregate AI R&D progress by a factor of two. The filing does not spell out which safeguards follow from crossing the threshold, and the company has published no separate statement on it.

Independent evaluator METR reported on Nov 5 that Anthropic's models were improving at about twice their earlier rate between March 2024 and October 2025, with about 80% of the acceleration attributable to AI taking part in its own development, and said the threshold may already have been crossed before July 15; Anthropic redacted most of the most informative evidence in that report, and the roughly 100-page full report has not been published. The filing also says the company does not currently expect to safely use Claude to autonomously design and build successive generations of advanced models without meaningful human supervision.

Key facts

1. Filing: Second amended prospectus filed Nov 10, 2026; the previous version said the threshold "may imminently" be met, the new one says it has been met
2. Threshold: Triggered when AI's contribution to its own development approximately doubles the rate of model improvement, requiring stronger safeguards
3. Basis: Internal estimates for future models and the overall rate of progress, not observed performance of current models
4. METR comparison: METR's Nov 5 report put improvement at about twice the earlier rate, with about 80% of the acceleration attributable to AI in its own development
5. Not disclosed: Which safeguards follow has not been spelled out; the roughly 100-page full report has not been published