Anthropic says future models may soon meet its AI acceleration rule

Anthropic told the U.S. Securities and Exchange Commission on November 2, 2026 that its internal estimates and expectations for future models indicate it may soon meet its own rule for “dramatically accelerated progress attributable to AI.” In the same Form S-1/A, the company says it has not designated any current model as over that line, and that it does not expect to let Claude design successor frontier models fully on its own without meaningful human supervision.

What did Anthropic tell the SEC?

The document is an amended registration statement, Form S-1/A, for Anthropic, PBC. The EDGAR index lists a filing date of November 2, 2026 and an acceptance time of 09:24:24 that day, under accession number 0001628280-26-068676.

In the risk factors, Anthropic writes that using its own systems to speed research “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 says much of its evidence on that pace comes from lagging indicators, and that it has in the past failed to identify a significant acceleration until after it had occurred, and may do so again.

The filing then states:

For example, 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.” We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data.

It adds that it is “evaluating the implications of meeting the AI research and development capability threshold” and “will release further reports in the future as we set forth in our Responsible Scaling Policy.”

Elsewhere, describing what it calls “closing the loop,” Anthropic says each version of Claude assists in building the next. The work it lists includes running experiments, generating and processing training data, automating evaluations and safety audits, and writing and reviewing code. It also says that using AI to build more capable AI “could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress,” and that its Responsible Scaling Policy covers the approach to “what has been termed ‘recursive self-improvement.’”

The same passage sets a limit in Anthropic’s own words: “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.” It says it is considering updates to the policy “to keep humans in control of frontier model development at Anthropic.”

The filing is also the company’s public account of its recent size. Anthropic states that revenue was $17.0 billion for the three months ended September 30, 2026, up from $1.4 billion for the three months ended September 30, 2025. For the 2026 quarter it states an operating margin of (3)% and an operating loss of $567 million, compared with an operating margin of (224)% and an operating loss of $3.1 billion in the year-earlier quarter. Those figures are the company’s statements in the registration document.

Julie Bort’s report for TechCrunch, published the same day, highlighted the recursive-self-improvement language. The sentences above are the wording in the filing itself.

What does “dramatically accelerated progress” mean in the policy?

The phrase in the filing points at Anthropic’s Responsible Scaling Policy, not at a new public benchmark. The current policy text is version 3.4, effective July 8, 2026. Anthropic’s policy page, which says it was last updated August 14, 2026, still lists version 3.4 as the current policy. The August 14 note on that page is a risk report, and the page does not list a policy version after 3.4.

Version 3.4 says the automated R&D threshold is met if Anthropic determines either of two things. The first is that its 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).” The second is that there is “dramatic acceleration” of the pace of AI progress for reasons that likely relate to automating AI R&D.

For that second case, the policy says Anthropic would look for a situation where it observes or expects “double the rate of progress” in “AI aggregate capabilities,” compared with both the rate it would expect and the fastest rate of extended progress it has observed without significant AI contributions to AI R&D. A footnote says that extended-progress comparison is “over at least three model generations.” The doubling also has to be plausibly “substantially attributable to the automation of research and/or engineering,” as opposed to other factors “such as increased headcount, compute, or general productivity,” and the continuation of the trend has to seem likely to lead to even greater acceleration.

The policy’s footnote defines the doubling this way: “‘Double the rate of progress’ means ‘as much progress in one year as one would see in two years at baseline.’” The example in that footnote: if baseline progress involved a 3× scaleup in compute and a 3× improvement in algorithmic efficiency, for a 9× “effective scaleup,” then double the rate of progress “would entail something like an 81× effective scaleup.” The footnote adds: “This is not the same idea as ‘doubling researchers’ productivity,’ since doubling inputs does not necessarily double the rate of progress.”

The July 8, 2026 changelog says this threshold “is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize.” It also records a clarification: Anthropic would not treat the threshold as crossed if the overall rate of AI progress were constant or slowing, even if that rate were dramatically faster than it would have been without advanced AI tools. In that case, the changelog says, there could be a strong argument for expecting the trend to continue rather than accelerate.

A glossary entry on this site is the right scale for reading any later scorecard: a named rule with a date, not a single percentage on a public leaderboard. Anthropic has not published, in this filing, the internal estimates behind the “may imminently meet” sentence.

Has Claude already crossed that line?

The filing’s answer on current models is direct: “We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data.”

The “may imminently meet” clause is attached to “internal estimates and our expectations regarding future models, rather than our current models.” Readers who want the distinction in one place can hold both sentences together. Anthropic is describing a forecast about models it has not designated as over the line, and it says its view of the pace can lag the event.

That is a different claim from a product announcement that a released Claude model now trains its replacement alone. The supervision sentence quoted above is the company’s statement on that point.

Why does it matter for AI agents?

An agent, on this site, is a model plus a harness: tools, a loop, permissions, and a stop. Autonomy is how far that loop may run before a person confirms the next step. Anthropic’s filing is about where that pattern sits inside a lab that builds the language model itself.

The practical example Anthropic gives is Claude Code, which it calls “our agentic coding product.” It says that as models advanced, Claude’s success rate on challenging, open-ended engineering problems that lack a clear specification reached approximately 90% in September 2026, up approximately 50 percentage points in nine months, “based on our internal benchmarks.” It says the average daily volume of code contributed to its codebase per engineer in the three months ended September 30, 2026 was over 45 times the average daily volume in the year ended December 31, 2024, and that over 90% of contributions to the codebase in September 2026 were written by Claude. Those are Anthropic’s figures, from internal measurement, published in the registration statement. They are not an independent audit.

The filing describes the shift in how people work with that system: engineers “increasingly direct and review its work rather than write code line by line,” after Claude Code “progressed from suggesting code to writing and running it directly and working autonomously over long time horizons.” That is tool use inside a company, with a person still in the review role the filing describes.

The policy threshold is a further step. It is about whether automated research changes the pace of AI progress in aggregate, not only whether a coding agent drafts a pull request. The filing treats that step as a safety risk the company says it is still measuring, and as a reason it may delay or limit models. It points to an earlier case: it says past precautionary measures included delaying releases and limiting initial availability, “such as with Mythos through Project Glasswing,” and that future measures could affect whether, how, and when new models or capabilities ship.

For a plain-language reading, start with how agents work and the Everyone door. For the harness and task questions, see harnesses, task performance, and the Builders page. The research framing of the policy lives with the research topic and the Research audience page. The public pattern of interleaving a model’s reasoning with actions is older than this filing; the site’s notes on ReAct cover that 2022 paper.

What happens next?

Anthropic says it will release further reports as the Responsible Scaling Policy requires, and that it is considering policy updates so humans stay in control of frontier model development. Version 3.4 says the company will publish a Risk Report every 3–6 months. It also says it will publish an analysis of an individual model’s relevant risks within 30 days of determining that an internally deployed model could pose high-stakes misalignment or automated R&D risks that significantly exceed those of all models for which it has already analyzed such risks in public.

The filing also says Anthropic cannot predict precisely when it will decide extra safety measures are necessary, what those measures will be, or how long they will stay in place. It says such measures could become more frequent and more costly as capabilities advance, and could slow its capability gains relative to competitors that do not adopt similar constraints.

The November 2 document names no new Claude model, states no new API price, and gives no date on which the threshold will be declared met. Until Anthropic publishes the further report it says the policy calls for, the sourced claim is the one in the risk factors: estimates for future models may meet the definition soon, and no current model has been designated as past it.

This filing is one item on the daily news beat. The Claude directory listing stays a product pointer; confirm current plans and model names on Anthropic’s site.

Frequently asked questions

Did Anthropic say a current Claude model has crossed the threshold?

The Form S-1/A says Anthropic has not designated any current model as exceeding the threshold, and that it continues to assess the underlying data. The “may imminently meet” wording is about internal estimates and expectations for future models.

What does “dramatically accelerated progress attributable to AI” refer to?

In the filing, that is Anthropic’s name for a definition in its Responsible Scaling Policy. Version 3.4 treats “dramatic acceleration” as one of two ways to meet the automated AI R&D threshold. The doubling it describes is aggregate AI capability — as much progress in one year as two years at baseline — judged plausibly due to automating research or engineering. The policy’s footnote says that is separate from doubling researchers’ productivity.

Is this the same as Claude coding without a person in the loop?

The filing separates the two. It describes Claude Code writing and running code on long tasks, with engineers directing and reviewing. It also says Anthropic does not anticipate safely using Claude to fully autonomously design and build successive frontier models without meaningful human supervision or understanding, if and when that becomes technically possible.

What revenue figure does the filing report?

Anthropic states revenue of $17.0 billion for the three months ended September 30, 2026, compared with $1.4 billion for the three months ended September 30, 2025. It states an operating loss of $567 million for the 2026 quarter.

Where is the primary source?

The amended registration statement is the Form S-1/A on EDGAR. The rule it cites is Responsible Scaling Policy version 3.4, linked from Anthropic’s policy page.