Executive Summary

Key Figures

Four numbers come with this story. The first two say how much of the work inside Anthropic is already being done by AI. The other two say how large the yardstick for crossing the threshold is, and how rarely the determination that uses it reaches the public.

Source: Anthropic S-1/A (SEC EDGAR, 2026-11-02), Responsible Scaling Policy version 3.4.

Anthropic filed its registration statement for a Nasdaq listing on October 26, 2026, and posted the first amendment to it a week later, on November 2. The registration number is 333-299633 and the accession number is 0001628280-26-068676. Both documents sit in full on the Securities and Exchange Commission's electronic filing system, so the same passage can be set side by side and matched word for word.

The passage that changed is the risk factor about using the company's own AI to build the company's own AI. One sentence that stood in the original was rewritten in the amendment, and three sentences were added after it. The diagram below sets that spot out, one version above the other.

The amended sentence reads: "As a result, 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 same place in the October 26 original read: "As a result, we may not identify a significant acceleration in our rate of development until after it has occurred." One clause is the whole of the difference. To the possibility of missing an acceleration ahead, the amendment adds the fact of having missed one before.

The sentence added after it is the one the coverage picked up: "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.'" Two more sentences follow. One says that no current model has been designated as exceeding this threshold and that the company continues to assess the underlying data. The other says it is evaluating what meeting the AI research and development capability threshold would imply, and that it will issue a further report as the policy provides.

That was not the only spot to change on the same day. The amendment revised three more, and all three point the same way. One is the sentence about a pace fast enough to set safeguards in motion. The original said model capabilities could advance at such a rate; the November 2 text says they may have already advanced, and may continue to advance, at such a rate. The second is a sentence that was not there before. As models take on more of the research and development work, the company says it is weighing changes to the policy that would require human review and approval of certain decisions in that process.

The third sits in the business description rather than the risk factors. Anthropic describes the arrangement in which Claude helps build the next generation of models as closing the loop, and presents it as a structural advantage of the company. Right after that description the revised text adds three sentences. It believes its investment in safety techniques lets it manage this risk. 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. And it is weighing updates to the policy that would keep humans in control of frontier model development at Anthropic.

The threshold behind that sentence lives in the voluntary rules Anthropic wrote for itself, the Responsible Scaling Policy. Version 3.4, in force since July 8, 2026, is the current one, and the item inside it on automated research and development in key domains is where the threshold is defined. If either of two branches holds, the threshold counts as met.

The first branch is the replacement of people. It is the case where the company's models are judged able to fully substitute for all of Anthropic's research scientists and research engineers at competitive cost, specifically "within a factor of 5." The second branch is the one the new risk factor sentence invokes: a dramatic acceleration in the rate of AI progress for reasons tied to the automation of AI research and development.

Two conditions hang on the second branch. First, the rate of progress in overall AI capability has to be observed or expected to double against two baselines at once. One baseline is the pace the company would otherwise have expected. The other is the fastest sustained rate of progress observed in a period when AI was not contributing much to AI research and development, where sustained means spanning at least three model generations. Then comes the second condition. It has to be plausible that the doubling is substantially attributable to the automation of research or engineering rather than to added staff, added compute, or general productivity gains, and the acceleration has to look likely to grow further if the trend holds.

How large a number a doubled rate really is, a footnote spells out. A doubled rate means delivering within one year what the baseline would have delivered in two. The example attached to it is sharp. If the baseline has compute growing 3x in a year and algorithmic efficiency improving 3x, for an effective scale-up of 9x, then a doubled rate corresponds to an effective scale-up of 81x. The same footnote is firm that this is not the same thing as doubling researcher productivity, because doubling the inputs does not double the rate of progress.

The Risk Report Anthropic published in August says in one line what crossing that line is meant to signal. The doubling is built to work as an early warning, not as evidence that the threat has already arrived. That same passage also names the sharpest form of this threat model: a regime where automation compounds and effective scale jumps by 10³ to 10¹⁰ within a single year. The doubling line is drawn to ring well before that regime begins.

One thing is worth noting. The phrase the amendment puts in single quotes, 'dramatically accelerated progress attributable to AI,' does not appear in that form in version 3.4. The words the policy uses are "dramatic acceleration," and "attributable" appears once, in the second condition about automation. The sense does not diverge, but the name the filing quotes as the policy's definition cannot be found as such in the policy document.

That this threshold is awkward to work with is not an outside assessment. Anthropic wrote it in the change log at the back of the policy document. The first of the five changes version 3.4 made is this automated research and development threshold, and the note on it closes this way: "This threshold is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize."

What proved difficult is recorded in the same note. The issue version 3.4 settled was a single case. Suppose the overall rate of AI progress is flat or even slowing, but is estimated to have been far slower still without advanced AI tools. Should that count as crossing the threshold? The answer written there is no. In that situation, the more convincing account is that deceleration from non-AI factors keeps offsetting the acceleration from AI, so the trend carries on rather than steepens.

What this one paragraph describes is the nature of the determination. Telling whether the threshold has been crossed takes more than measuring the actual rate. It needs an estimate of what the rate would have been without AI. The branch that did not happen cannot be observed, so an estimate enters, and where an estimate enters, judgment enters. Version 3.4 rewrote this passage in July 2026 because internal discussion had shown that the same situation could yield different conclusions.

The amendment does not set out the procedure for the determination. It says only that the company will issue a further report as its policy provides. That procedure sits in chapter 3 of version 3.4. The document that carries the determination is the Risk Report, and the publication interval and the sign-off path are written down with it.

4.1Cadence

Risk Reports come out every 3 to 6 months. The policy is explicit that, unlike a system card, a report does not appear with every model release. What a report covers reaches back no further than 30 days before its publication date. There is also a case that falls outside the cycle. If a model deployed internally is judged to carry far greater risk, in high-risk misalignment or in automated research and development, than any model whose risks have been publicly analyzed so far, an analysis of that model has to be published within 30 days of that judgment. A separate clause puts models deployed internally for autonomous large-scale research inside a report's scope at a minimum.

These reports do not stay inside the company's own files. They are actually published, and two have appeared so far. The first went up on February 24, 2026, and the second on August 14, both as public versions that mark where material has been withheld. The August report's coverage date is July 15, and it says it covers the period from the first report on February 24 up to that day. The determination this article turns to later sits inside it.

4.2Sign-off

There are four steps. Internal experts run the risk assessment and draft the report. Broad internal comment follows, aimed at weaknesses in the methodology and gaps in the analysis, and outside experts are usually asked for comment as well. The third step is the decision. The report and the comments go to the CEO and the Responsible Scaling Officer, and in the policy's own terms those two make the final determination on the adequacy of the risk assessment and on subsequent deployment or development plans. The officer is a designated member of staff charged with carrying out the policy, and judgments on how the policy is interpreted and applied fall inside that role. The fourth step is notification. After approval, the two promptly share the decision, the report, and the internal comments with both the Board of Directors and the Long-Term Benefit Trust.

The order matters. The Board and the Long-Term Benefit Trust do not take part in the determination. They receive it once it has been made. There is one exception. Where the argument that the company's own added risk is relatively small, given risks other developers have already created, carries significant weight in the decision, explicit approval from the Board and the Trust is required instead of a call by the two officers.

4.3Outside Eyes

Outside eyes are not absent. They are not always there either. Version 3.4 says it will work toward a practice of comprehensive, public external review of Risk Reports, and adds that no well-established organizations or procedures for this kind of practice exist yet, so it is approaching the matter as an experiment. Review becomes mandatory only when two conditions hold together: the report covers models that are "highly capable," and the report is at the same time "significantly redacted." The second condition means that enough material has been withheld for a reader of the public version to be unable to judge whether to agree with the conclusion.

The definition of the first condition lands exactly on this article's subject. A model counts as highly capable, under the policy's definition, once Anthropic concludes that it crosses the threshold for automated AI research and development. That is the very threshold the S-1/A calls imminent. So the moment a model crosses it, one of the two switches that turn the obligation on has been flipped, and if the report is significantly redacted as well, then for the first time the company is required to show an unredacted report to at least one external reviewer and have that reviewer write a public report. That reviewer, by the policy's terms, faces no restriction on what to write apart from confidential information, and is not barred from writing about concerns with the report or with Anthropic's conduct.

So far the obligation has never been triggered. The August report writes that state of affairs down plainly. Since version 3.2, on April 29, 2026, gave the Long-Term Benefit Trust the power to request an external review and to approve the choice of reviewer, the Trust has not requested one and the policy has not required one, though pilot external reviews have continued. METR took the automated research and development section of the February report, and SecureBio took the chemical and biological risk sections.

The review METR published on May 8 is the one document that answers this article's question from outside. Its conclusion splits in two. It agrees with the report's bottom line, that the risk of a catastrophe from the model of the time automating research and development in any domain is very low. It then says the evidence the report presents cannot carry that conclusion: "we think the evidence presented in the report is inadequate to establish this." Two recommendations it set down alongside stand out now. One is that the overall argument misses the possibility of substantial AI research and development acceleration before its full automation. The other is that Anthropic report other sources of evidence that could serve as more leading indicators of AI progress.

Put together, what external review attaches to is a finished Risk Report, and even that has yet to be required. The final determination on whether the threshold has been crossed stays with two people.

So what did the last published determination on this threshold say? Section 3.5 of the August report is the place. The section is titled "Acceleration of AI progress due to automated AI R&D," and it takes up exactly the second of those two branches. Earlier in the report the company writes that its models as of the coverage date meet neither branch, and section 3.5 lays out the grounds for the second. The conclusion is that the overall pace of progress remains below the policy's doubling line.

That sentence does not stop there. It goes on to say that even with the additional sources of data the conclusion is uncertain and subject to some lag, "such that we would have difficulty measuring very recent acceleration." The qualifier is attached to a determination whose coverage date is July 15, and three and a half months later the same company wrote in its listing document that the threshold may be imminent.

What the measuring was done with is on the record too. The main evidence presented for the acceleration criterion has been trends in the Anthropic ECI, an internal fork of Epoch AI's Epoch Capability Index rebuilt to run on in-house benchmark results, and the August report is clear that these cannot be compared directly with scores on Epoch AI's public leaderboard. The report says Anthropic watches in-house leading indicators that may point ahead of later index trends as well, and the very next sentence states that what those indicators are and how they are moving is sensitive and not included in the public version. That is the exact spot METR had recommended strengthening three months earlier.

That is as far as the dashboard visible from outside goes. The trend of one index, two determinations in February and August, and a note that the most recent stretch is hard to see. What the November 2 filing calls imminent did not come off that dashboard. By its own account, it comes from internal estimates and expectations about future models rather than current ones.

Read as nothing more than news that AI is building AI, this language leaves little behind. The more useful reading takes it as a confession about measurement. Set down as Anthropic wrote it: the evidence about the pace of acceleration comes largely from lagging indicators, and the ability to forecast or measure that pace, including acceleration attributable to recent or unreleased models, is limited. As a result of that limit, the company has in the past noticed an acceleration only once it had happened.

Why watching your own system is hard is all in there. The threshold is written as a number, but filling that number takes three things already in place: a metric that decides what counts as progress, a record of the past stretch to compare against, and an estimate of what would have happened without AI. Build any of the three after the event and it is too late. To say the evidence comes from lagging indicators is to say that no dashboard was running while the event was happening.

The previous section is what that difficulty looks like in the flesh. A dashboard is not entirely missing. It is read once every six months, the most recent determination wrote down that it cannot see the recent stretch well, and the leading indicators that would shore up that determination were left out of the public version as sensitive. That the measurement runs after the fact is already written down as numbers, in a publication interval and a coverage date.

That AI's share of the work inside Anthropic is not small is something the same filing shows in figures. Claude wrote more than 90% of the contributions to the company's own codebase in September 2026, and daily code contribution per engineer in the third quarter of 2026 ran more than 45 times the daily average across 2024. These figures are not evidence that the threshold has been met. What version 3.4 set out to measure is the rate of progress in overall AI capability rather than lines of code or productivity, and the footnote makes a point of keeping added input and an increased rate apart. The two figures show how far automation has come inside the company. They are not inputs to the determination.

Here is what there is to take away for people who work with data. What changed can only be confirmed afterward from what was recorded while the change was passing through. Without a record of when the baseline was taken, how often the measurement was run, and who approved that measurement and when, there is no way to account for the change. If one of the most measurement-literate companies in the world wrote in a document for investors that its ability to measure how fast its own capabilities are rising is limited, then in an organization with a shallower measurement apparatus the same thing happens more quietly.

This is what Pebblous keeps coming back to in talking about AI-Ready Data. The worth of data comes from the record of where it came from and under what conditions, not from how much of it there is. This filing shows that the point binds the side that builds models as much as the side that uses them. To say when your own system changed, you first need what you measured before it changed.

Thank you for reading this far. The filing quotations in this article come from a direct comparison of the S-1/A with the October 26 original, and the threshold definition and the procedure can be checked in version 3.4 of the Responsible Scaling Policy. Count how many of the models your team runs you could give a date for, the day their behavior changed, and we would be glad to hear the number.

Primary Sources

• 1.Anthropic, PBC. (2026). "Amendment No. 1 to Form S-1." SEC EDGAR, registration no. 333-299633, 2026-11-02. (Source for the added risk factor sentence and the added business description text)
• 2.Anthropic, PBC. (2026). "Form S-1." SEC EDGAR, 2026-10-26. (The original used as the point of comparison)
• 3.Anthropic. (2026). "Responsible Scaling Policy, Version 3.4." Effective 2026-07-08. (Threshold definition, the footnote on a doubled rate, Risk Report procedure, change log)
• 4.SEC EDGAR. (2026). "Anthropic, PBC S-1/A filing index." Accession no. 0001628280-26-068676, 2026-11-02. (Confirms the filing date and the registration number)
• 5.Anthropic. (2026). "Risk Report: August 2026 (Redacted)." Published 2026-08-14, coverage date 2026-07-15. (Section 3.5 acceleration determination, Anthropic ECI, leading indicators withheld, state of external review)
• 6.Anthropic. (2026). "Risk Report: February 2026 (Redacted)." 2026-02-24. (The first published Risk Report, the one METR reviewed)

External Review

• 7.Jurkovic, Nikola, Beth Barnes, and Hjalmar Wijk. (2026). "Review of the "Risks from automated R&D" section in the Anthropic Risk Report (February 2026)." METR, 2026-05-08.
• 8.SecureBio. (2026). "Review of Anthropic's unredacted chemical and biological risk sections." (Pilot review of the chemical and biological sections of the February report)

Press

• 9.Bort, Julie. (2026). "Anthropic warns that it's very close to RSI capability, aka self-building AI." TechCrunch, 2026-11-02.
• 10.Gizmodo. (2026). "Anthropic Claims to Be Closing in on Recursive Self-Improvement." 2026-11.