Anthropic, the company that makes the Claude AI models, says it may soon reach a point its own safety policy sets aside for extra precautions: AI dramatically speeding up the pace of AI progress itself. In an updated version of its filing to sell shares to the public, sent to the Securities and Exchange Commission on Monday, Nov. 2, it added that its internal estimates and its expectations for future models (not its current ones) "indicate that we may imminently meet" its safety policy's definition of "dramatically accelerated progress attributable to AI." It says it has not judged any of its current models to be past that line, and that it is still assessing the data.
The line comes from Anthropic's Responsible Scaling Policy, the published rules it sets itself for handling the risks of its most capable models. The policy counts the line as crossed if Anthropic determines that its models could fully replace all of the company's research scientists and research engineers at no more than five times the cost. It is also crossed if Anthropic sees or expects AI to be advancing twice as fast as both the pace it would expect and the fastest stretch it has seen without much help from AI, as much progress in one year as would normally take two, that speed-up plausibly comes in large part from AI doing research and engineering work rather than from more staff or computing power, and the trend looks likely to speed up even more. The filing's wording points to that second test. The policy says the line is meant to catch the start of dramatic recursive self-improvement: an AI improving itself, then using the improved version to improve itself again, round after round.
Claude already helps build Anthropic's next models. The filing says many parts of the research, training and testing of those models now involve Claude, including running experiments, making and processing training data, running tests and safety checks, and writing and reviewing code, and that more than 90% of the contributions to Anthropic's code in September 2026 were written by Claude. Anthropic calls this "closing the loop" and presents it to investors as a strength.
The update also changes a warning. The first version of the filing, sent Oct. 26, said Anthropic "may not identify" a big speed-up in its pace of development until after it had happened. Both versions put that down to evidence that mostly comes from figures that show change only after the fact. The new one adds that, in the past, it has not spotted such a speed-up until afterward. It adds that it does not expect to be able to safely let Claude design and build later generations of its most advanced models on its own, "without meaningful human supervision or understanding," if that ever becomes technically possible, and that it is considering changes to its policy "to keep humans in control."
Under the policy, when Anthropic publishes a report on the risks of a model past the line and leaves out details an outside safety expert would consider important for judging the risk, the report must go through a full review by at least one outside reviewer. If Anthropic has, or is about to have, such a model and clear evidence that no rival will soon build one, the policy says it will need a strong case that the risk of catastrophe is contained, and will delay development and release as needed to make that case, unless it no longer believes it has a big lead. The filing says earlier precautions included delaying the release of models and limiting who could first use them, and that such steps could slow its progress compared with rivals that don't take them.
The filing does not say which future model the estimate is about, when, beyond "imminently," the line might be crossed, or what figures lie behind it, and it does not say what the policy changes would be, beyond keeping humans in control. Anthropic's published safety plan has said it is plausible that, as soon as early 2027, its AI could fully automate or dramatically speed up the work of large teams of top researchers in fields including AI itself. Anthropic says it is weighing what meeting the threshold would mean and will release further reports as its policy requires.