Anthropic tells investors it may soon reach its own threshold for AI-driven acceleration

Anthropic added new language to the Risk Factors section of its revised S-1/A, filed on Monday, November 2, 2026. It warns that the company may soon meet a threshold in its Responsible Scaling Policy: "dramatically accelerated progress attributable to AI." The filing ties that threshold to what has been termed "recursive self-improvement" (RSI). In the filing, RSI is a step toward AI that can design and build its own successors.

The statement matters because of where it appears. A registration statement is a legal document. Companies use it to tell investors what could go wrong, and they can be held liable for what it says. Here, a frontier lab told investors that its own research may be speeding up faster than its safety work can keep pace. It also says it might not notice that speed-up until after it has happened.

What the filing says

TechCrunch reported the passages, and its account is the source for the quotes below. The filing says Anthropic uses AI in its own research and development. It warns that "using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress."

The new risk factor goes further. According to the filing, Anthropic's "increasing use of AI systems to accelerate our own research and development may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures." It adds: "we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred."

On timing, the filing says Anthropic's "internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet" the policy's acceleration threshold. It also says: "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."

Anthropic also says what it will not do. "We do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible." The company says it is considering updates to the Responsible Scaling Policy "to keep humans in control of frontier model development at Anthropic."

TechCrunch notes that an S-1 is both a disclosure and a marketing document. Claims of being close to self-building AI could be read either way.

The capabilities behind the warning

The filing does not list the specific capabilities. An earlier Anthropic essay, "When AI builds itself", by Marina Favaro and Jack Clark, gives the clearest public account of them. It defines recursive self-improvement as an AI system "capable of fully autonomously designing and developing its own successor". It says this has not been reached.

The essay reports several measurements from inside the company:

• As of May 2026, "more than 80% of the code we merge into Anthropic's codebase was authored by Claude." A footnote says leadership has estimated 90% or more if scripts and experimental code are included.
• The typical engineer was merging eight times as much code per day as in 2024. The authors call this "almost certainly an overstatement" of the real productivity gain, because lines of code is an imperfect measure.
• "Claude can already match or outperform skilled humans at executing a well-specified experiment."
• In one AI safety project, agents ran the full research loop and recovered 97% of a performance gap. Two human researchers recovered 23%. Humans still chose the problem and wrote the scoring rubric.
• Anthropic picked moments in research sessions where the human's choice of next step had room to improve. At those moments, the best model's choice beat the human's 51% of the time in November 2025 and 64% in April 2026.

The essay also names what is missing. It says "large performance gaps persist" in choosing research goals, and calls this the gap between today's AI and a system that could design its own successor. It says the length of tasks AI can complete reliably has been doubling "roughly every four months." It projects that tasks taking a skilled person days could come into range this year, and tasks taking weeks in 2027. The authors also write that "recursive self-improvement is not inevitable."

Pause, policy and context

In the same essay, Anthropic says it would be good for the world to have "the option to slow or temporarily pause frontier AI development." It says it would likely pause if other frontier developers did so "in a verifiable manner." It argues that a unilateral pause "accomplishes much less". It also says that detecting covert training runs is harder than detecting comparable activity with other technologies.

Fortune reported in June that Anthropic had dropped a central commitment from its Responsible Scaling Policy in February. That is the same policy the filing now relies on to define the acceleration threshold. Fortune also noted that some observers questioned whether the timing of the pause proposal, shortly after Anthropic's confidential IPO filing, was partly about publicity.

What it means for work

The work being automated first is the work of AI researchers and engineers themselves. By Anthropic's own figures, most of the code merged at the company is already written by Claude, and well-specified experiments can be handed to it. The part that remains human, for now, is choosing which problems matter and deciding how to judge the results. Anthropic says that is where its models still trail. Its stated plan to keep "meaningful human supervision or understanding" means people will spend less time writing and running research. More of their time will go to reviewing, validating and approving what AI systems produce. The filing raises a question it does not answer: whether that oversight can keep up, when the company says it may only recognise an acceleration after it has occurred.