Anthropic says it may soon reach AI that speeds up its own development

Its filing says future models could cross a self-acceleration threshold in its safety policy, though no current model has.

Anthropic's amended S-1 filing, released Monday, includes a warning that the company may soon reach recursive self-improvement, the stage where AI becomes capable enough to accelerate its own training and development. The company says it already uses AI in its research and development and has in the past failed to notice major capability jumps until after they happened. It states that its internal estimates for future models, though not its current ones, suggest it may imminently meet the threshold for 'dramatically accelerated progress attributable to AI' set by its Responsible Scaling Policy. Anthropic says no current model has crossed that line and that it does not expect to let Claude design and build successive frontier models without meaningful human oversight, while it considers changes to that policy. Commentators note that an S-1 doubles as a marketing document, but the disclosure has drawn attention given Claude's recent gains in coding, math and cybersecurity.