TL;DR
- In its November 10 amended S-1 filing with the SEC, Anthropic wrote: "We made a determination that we have met this threshold." The threshold is from Anthropic's Responsible Scaling Policy — the company's own rulebook for when stronger safeguards apply. The specific trigger: AI automating enough of Anthropic's research that the rate of AI capabilities progress roughly doubles (two years of progress at the 2018-2024 baseline rate compressed into one year). Anthropic based the determination "on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models." No current deployed model has been designated as exceeding the threshold. Crossing it triggers additional safety requirements. - The sequence: November 3 amendment said Anthropic "may imminently meet" the threshold. November 10 amendment said it has met it. November 9: Claude Fable 5.2 released with 1,000-agent Fermat's Last Theorem formalization in 30 hours. The IPO roadshow launched November 10. Shares expected to begin trading November 19-20. In an IPO document filed with the SEC, Anthropic disclosed that its AI has crossed an internal accelerated-progress safety threshold. - Independent confirmation: METR, the AI evaluation nonprofit that Anthropic gave employee-level access to in October, estimated that Anthropic's models have been improving at approximately twice the baseline rate since October 2025. METR separately judged that Anthropic's August assessment (that it was below the threshold) was "overstated" and found it plausible the threshold was crossed as early as mid-July 2026. - What crossing the threshold does not mean: Anthropic is not pausing, not planning to pause, and does not expect to let Claude "fully autonomously design and build successive generations of frontier models without meaningful human supervision." The company is weighing unspecified policy changes "to keep humans in control" and acknowledged its "organization may lose visibility across certain processes" as it relies more heavily on Claude. - Separately: the US and eight allies signed the San Francisco Declaration (Pax Silica) on November 15, committing to a joint registry of AI technologies and supply chains vulnerable to Chinese interference. Signatories: US, Japan, Australia, Israel, UK, South Korea, India, Qatar, UAE. Germany, Sweden, Netherlands, and the EU Commission expected to sign next week. Timed three days before the Trump-Xi APEC meeting in Shenzhen.
Anthropic is about to become a public company. It is about to do so after disclosing in a legally binding securities filing that its AI is accelerating its own research at a pace the company's safety policy identified as requiring heightened safeguards.
This is a first in the history of publicly offered companies.
What the threshold is and what crossing it means
Anthropic's Responsible Scaling Policy is a document that Anthropic published and committed to follow as a governance mechanism. It defines specific "AI Safety Levels" (ASLs) and triggers that require specific behaviors as models become more capable.
The "Automated R&D in key domains" threshold is triggered when Anthropic's AI meets either of two conditions:
Condition 1: Anthropic's models can replace all of Anthropic's research scientists and research engineers, at no more than five times the cost of employing those teams.
Condition 2: Anthropic observes or expects a doubling of the rate of AI capabilities progress — specifically, two years of progress at the 2018-2024 baseline rate achieved in one year — and it is plausible that AI automating research is substantially responsible for the doubling.
Anthropic's November 10 disclosure triggered Condition 2. The company based its determination on its "overall rate of progress" and "what we expect from future models" — not on current deployed models alone. AI "has already been accelerating our research and development," the filing states.
What crossing the threshold requires: the RSP specifies that Anthropic must implement stronger safeguards. The specific requirements are not publicly detailed in the filing, but the threshold is designed to be the trigger for a qualitatively different safety posture — additional evaluations, more constrained deployment conditions, and processes to ensure human oversight of automated research contributions.
What it does not require: a pause, a shutdown, or ceasing AI-assisted research. Anthropic is continuing to operate, continuing to deploy models, and continuing the IPO. The company characterizes crossing the threshold as a governance event that triggers additional precautions, not an operational stop signal.
The specific text of the RSP change between filings is telling. November 3: Anthropic "may imminently meet" the threshold. November 10 (the same filing that set the IPO price range): "we have made a determination that we have met this threshold." Eight days, one week before IPO, Anthropic's governance assessment changed from "imminently approaching" to "crossed."
METR's independent assessment
METR — the AI safety evaluation nonprofit that Anthropic gave employee-level internal access to in October, to conduct independent risk assessments — examined Anthropic's own evaluation data. METR found that Anthropic's models have been improving on "well-specified tasks" at approximately twice the baseline rate since October 2025.
This is consistent with Anthropic's disclosure and extends it. METR's additional finding is more pointed: METR judged that Anthropic's August assessment — that it was still below the threshold — was "overstated," and found it plausible the threshold was crossed as early as mid-July 2026.
If METR is right that the threshold was crossed in mid-July, the sequence is: threshold crossed → Anthropic operated at this capability level for approximately three months before formally acknowledging it → the acknowledgment appeared in a securities filing timed to the IPO, not in a standalone public announcement.
The timing is not sinister — IPO filings are legally required to disclose material facts, and the November 10 amended S-1 is the appropriate mechanism for disclosing a material safety policy event. But the gap between plausible threshold-crossing (mid-July per METR) and formal acknowledgment (November 10) is worth noting. During that period, Fable 5.2 was developed, trained, and released — the model that formalized Fermat's Last Theorem in 30 hours and solved open problems in black hole physics.
What this disclosure means for the IPO
Anthropic is raising up to $90+ billion from public market investors at a valuation of approximately $1-$1.5 trillion, one week after disclosing in a securities filing that:
1. Its AI has crossed the threshold for dramatically accelerating its own research 2. The company may lose visibility across certain processes as it relies more on Claude 3. It is weighing policy changes to maintain human control 4. It does not expect to safely let Claude fully autonomously build its successors without meaningful human supervision
Each of these is a disclosed material risk. Investors reading the S-1 are informed of these facts and are nonetheless purchasing shares at the offered price. The public market is, by definition, pricing these risks into the valuation — either discounting them as manageable concerns consistent with strong revenue growth, or accepting them as intrinsic to the investment.
The disclosure also creates legal obligations that extend beyond the IPO. As a public company, Anthropic will be required to disclose material changes in its safety policy status in future filings. If the company determines it has crossed additional safety thresholds — or that the current threshold has implications it did not anticipate — those disclosures will be public.
This is the first time the market will price AI safety policy compliance as a disclosed material variable in a public company's SEC filings. How ANT trades when it opens will be the first direct market signal on that pricing.
The San Francisco Declaration: Pax Silica
Signed November 15 — three days before the Trump-Xi APEC meeting in Shenzhen — the San Francisco Declaration formalizes what US Under Secretary of State Jacob Helberg described as a historic commitment: treating AI supply chain security as a shared responsibility among allied nations.
The specific commitment is a joint registry: the nine signatories (US, Japan, Australia, Israel, UK, South Korea, India, Qatar, UAE) will jointly identify sensitive AI technologies and supply chains considered vulnerable to Chinese interference, share information on diversification strategies, and establish working groups on AI supply chain vulnerabilities.
The scope extends beyond semiconductors to include critical minerals and energy infrastructure. This reflects the full supply chain logic: a chip cannot be designed, manufactured, or shipped by any single country. The know-how to design is concentrated in US/EU companies. The manufacturing is concentrated in TSMC (Taiwan/Arizona). The critical minerals are concentrated in China. The machinery is concentrated in ASML (Netherlands). The accord is designed to treat this distributed dependency as a collective security concern.
Germany, Sweden, the Netherlands, and the EU Commission are expected to join next week. That would add the Netherlands (ASML — the only manufacturer of extreme ultraviolet lithography machines required for leading-edge chip fabrication) and Germany (ZEISS — provides the optical components for ASML's machines) to the registry commitment.
The timing is deliberate: signing three days before Trump-Xi at APEC creates leverage for the US in those discussions. China knows the nine-country registry will identify Chinese-influenced supply chain nodes before the summit; it cannot unknow that.
What the registry will specifically include is unknown — the State Department had not published the declaration text as of the signing date. The legal and trade implications depend on what "jointly identify vulnerable technologies" translates to in practice: coordinated export controls, shared audit requirements, investment screening, or some combination.
The week before ANT begins trading
November 18: IPO pricing decision (expected Thursday) November 19-20: Anthropic shares begin trading on Nasdaq as ANT November 18-19: Trump-Xi APEC meeting in Shenzhen; first formal US-China superintelligence dialogue meeting
The company going public has disclosed, in legally binding SEC filings, that its AI is now accelerating its own research at the rate its safety policy identified as requiring heightened precautions. The public market will determine on Thursday what that is worth.



