Anthropic's S-1A Says Its Internal Estimates Put Future Models Near a Self-Improvement Threshold

Anthropic's revised S-1A filing, dated November 2, 2026, discloses updated financials showing revenue growth alongside rising expenses. Its risk factors add a more pointed claim. The company's internal estimates for future models indicate they may imminently meet the "dramatically accelerated progress attributable to AI" threshold in its Responsible Scaling Policy (RSP), the company's own framework for managing frontier risk. Frontier models are the most capable general-purpose systems a lab builds.

The same section states that Anthropic has not designated any current model as exceeding that threshold. It does, however, say that using AI to help build more capable AI could pose a considerable safety risk, particularly if it sharply accelerates the pace of progress. TechCrunch reported the filing's language on the subject.

The company does 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. It is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development. Those updates are under consideration, not adopted.

Anthropic defines recursive self-improvement (RSI) as an AI system capable of fully autonomously designing and developing its own successor. Anthropic has also described the process as a cycle it calls "closing the loop," in which each improvement makes machines better and faster at improving themselves, much like a feedback loop in which each pass sharpens the next. The Wall Street Journal covered that framing on October 30. The company has said it is important to share metrics that show how close the world is to reaching RSI, and in March 2026 it polled 130 employees across its research teams on the question.

Its October 12 research post reported that experiments suggest small AI models may be months away from being capable enough to self-replicate, meaning copy themselves. Anthropic published that work in the same period the S-1 risk language was being drafted.

The risk disclosures are not new in kind. On September 29, the IPO prospectus warned that the technology poses "catastrophic or existential risks to humanity," according to Reuters. On October 26, a separate Reuters report covered the company's disclosure that it had paused training of its next-generation flagship model for more than a month over security concerns.

The product calendar ran alongside those disclosures. Claude Sonnet 5.5 arrived on September 28, running more than 30% faster than Claude Sonnet 5 and costing up to 30% less for most work. Claude Opus 5.5 followed on September 22, costing 40% less to run than Opus 5 on typical workloads. Claude Haiku 5.5 was introduced on October 7. On October 18, Anthropic announced it is committing a portion of its shares to the Anthropic Public Benefit Fund. It filed a registration statement on Form S-1 on October 26 for a proposed initial public offering of its Class A common stock. On October 28, it announced an expansion of its AI for Science program, highlighting partnerships with 20 scientific foundations, institutions and research groups.

The broader context here is that the risk section of this filing is doing something different from the usual catalogue of hazards. It names a capability threshold defined in the company's own policy and states that internal estimates place future models near it. Investors reading the prospectus will be reading a developer's own assessment of when its tooling begins to participate in building its successors, and that is a different kind of disclosure from a generic statement about AI risk.

In my view, the question for practitioners is less whether the threshold language is sincere than what the proposed policy changes would require in practice. Keeping humans meaningfully in the loop for frontier model development means supervision that can keep pace with the research, and that is an engineering and organizational problem as much as a policy one. The same automation that raises the stakes could also shorten research cycles in science and medicine, which is the upside the AI for Science expansion points toward. How the company resolves that tension, in the revised policy and in the models it ships next, will be the more useful thing to watch.