A warning filed as a risk factor
Most of Anthropic's amended prospectus is arithmetic. This part is not. In the Risk Factors section of the S-1/A filed on November 2, the company added language about what happens if AI starts speeding up the work of building AI. Using AI to help build more capable AI, the filing says, could pose a considerable safety risk if it sharply accelerates the pace of progress, and the area is one its Responsible Scaling Policy covers as the company approaches what has been termed recursive self-improvement.
Then the filing puts an estimate on it. Internal estimates and expectations regarding future models, rather than current models, indicate that Anthropic may imminently meet its policy definition of "dramatically accelerated progress attributable to AI." That sentence deserves to be read with the sentence that follows it in the same passage: the company has not designated any of its current models as exceeding the threshold, and it continues to assess the underlying data. This is a forecast about models that do not exist yet, measured against a threshold the company wrote itself. Both halves of that description matter.
What the threshold actually measures
The phrase sounds like a mood. In Anthropic's policy it is meant to be a measurement. Under the Responsible Scaling Policy v3.4, effective July 8, 2026, one route to the threshold is a doubling of overall capability progress attributable to AI, judged against the progress expected without AI. The doubling is in the capabilities of the resulting systems, not in how much code a researcher ships or how many experiments a team runs. Anthropic clarified exactly that distinction in an earlier revision of the policy, in April 2026, because the easier version of the test — researchers working twice as fast — would say little about whether the models themselves are compounding.
The distinction matters because faster research activity does not automatically mean faster progress in model capabilities. A team might run twice as many experiments without producing systems that improve twice as quickly. Anthropic's stated concern is the latter: AI contributing to an acceleration in the capabilities of the systems being developed, rather than simply making the researchers more productive.
That distinction is the useful part of the disclosure. A lab can double its research output and produce models that improve at the usual rate. The threshold Anthropic says it may approach is the stronger event: progress itself speeding up because AI is doing a material share of the work that produces progress. Nothing in the filing shows that event has happened. What the filing claims is narrower, that the company's own projections now place it close enough to describe as imminent in a document its lawyers had to sign off on.
The sentence that cuts against the forecast
Immediately before the forecast, the filing makes an admission that should be quoted whenever the forecast is. Anthropic's increasing use of AI to accelerate its research, it says, may cause model capabilities to advance more rapidly than its ability to develop, validate and implement safety, alignment and security measures. As a result, the company states, "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."
Put the two statements together and the problem becomes specific. Anthropic says its ability to forecast or measure capability progress is limited because the indicators it relies on are lagging. It is projecting that future models may imminently meet its acceleration threshold while acknowledging that significant acceleration may not be recognized until after it has happened. The filing does not say that the threshold has already been crossed. It says the company may be approaching it while its ability to detect the change remains imperfect.
The filing closes the topic with a control it intends to keep: Anthropic says it 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, if and when that becomes technically possible. It is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.
One qualification matters. An S-1 is a regulatory disclosure document, but it is also part of the material through which a company presents itself to prospective investors. Anthropic is disclosing a potential safety risk, not publishing an independent assessment of its own development trajectory. The underlying estimates are not included in the passage, and outsiders cannot reproduce the company's conclusion from the filing alone. The verifiable points are what Anthropic says it expects, what it says has not happened, and what limitations it acknowledges in measuring progress.
Sources
- [1] Anthropic, PBC, Amendment No. 1 to Form S-1 (Form S-1/A), U.S. Securities and Exchange Commission, filed November 2, 2026 (read first-hand — the Risk Factors passage stating that much of the company's evidence regarding the pace of AI-driven acceleration is derived from lagging indicators and that "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"; the statement that "our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI'" and that the company has "not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data"; and the Business section passage stating "Using AI to help build more capable AI could also pose a considerable safety risk" as the company approaches what has been termed "recursive self-improvement," and that "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" and is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development)Read source
- [2] Anthropic, Responsible Scaling Policy, Version 3.4, effective July 8, 2026 (official policy document, read first-hand — the automated R&D capability threshold is considered met if either models "would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs (i.e., within a factor of 5)" or there is "dramatic acceleration" of the pace of AI progress for reasons likely related to the automation of AI R&D, defined as observing or expecting "double the rate of progress in AI aggregate capabilities" compared to both the rate expected and the fastest rate observed in the absence of significant AI contributions to AI R&D, sustained over at least three model generations, where the doubling is plausibly substantially attributable to the automation of research and/or engineering; the document's footnote states this "is not the same idea as 'doubling researchers' productivity'")Read source
- [3] Anthropic, "Anthropic's Responsible Scaling Policy" (official policy page, read first-hand — version 3.4 listed as effective July 8, 2026; the July 8, 2026 changelog entry stating that the v3.4 update "revises our threshold for automated R&D to better track the threat model of concern"; the April 2, 2026 entry stating that v3.1 included "a clearer definition of the AI R&D capability threshold," clarifying that AI doubling the rate of progress means "doubling the rate of progress in aggregate AI capabilities" and not "doubling the productivity of researchers")Read source
- [4] TechCrunch, "Anthropic warns that it's very close to RSI capability, aka self-building AI," November 2, 2026, updated November 3, 2026 (the report that surfaced the new Risk Factors language; the filing passages quoted in this article were verified against the filing text itself, at the first source above)Read source