Anthropic Says It May Soon Cross Its AI-Builds-AI Line. Its Policy Ties the Brake to Rivals.

A safety threshold is only as strong as the action it forces. Anthropic told prospective shareholders this week that it may be about to cross the most consequential line in its own safety policy, the one meant to catch AI that speeds up the building of AI. Read the policy that defines the line, and the first independent measurement of how close Anthropic is, and a less comfortable picture appears: crossing it mainly triggers a review, and whether it triggers a delay depends largely on what OpenAI and Google have built by then.

The sentence Anthropic added to its IPO filing

On November 2, Anthropic filed Amendment No. 1 to its S-1, a week after the original registration statement that the site read for its revenue and channel concentration. The amendment adds this to its risk factors:

“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.”

The same paragraph concedes something more unusual for a securities filing: “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.” Translated out of legal register, Anthropic is saying its instruments lag, and that it has already been surprised once.

Two further passages in the amendment matter for what comes next. In the business section, Anthropic writes that 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,” and that it is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development.” In the risk factors, it says it recently “chose to delay pre-training a new frontier base model for more than a month” to harden security against weight theft, and that it is weighing rules “to require human review and approval of certain decisions in our research and development processes as our models perform an increasing share of that work.”

That last phrase is the one I would underline. The interesting word is “certain.” Every operations team that has automated a large share of its own change volume has had to decide which decisions keep a human signature. When automation started generating most of the routine configuration changes in a network operations group I ran at a large telecom, the change advisory board could not keep approving changes one by one; the volume made that theater. What worked was moving human approval up a level, onto the templates and the pipelines that produced the changes, and keeping individual sign-off only for changes that could not be rolled back. An AI lab writing “human review and approval of certain decisions” is facing the same design question at far higher stakes, and until it publishes which decisions, nobody outside can judge whether the gate sits on the pipeline or on the paperwork.

The Gadgeteer was among the first to compare the two filings line by line; the October version said only that Anthropic “may not identify” an acceleration, without the “in the past” admission or the “imminently” sentence.

What METR measured from inside the building

Three days later, METR published its independent assessment of AI R&D acceleration at Anthropic. It is the first output of the arrangement Anthropic announced on October 15 in an update on independent evaluators, which gave METR staff company hardware, office space and access to internal tools and systems. (METR notes its engagement started in mid-August and is distinct from the assessment announced that day; it also says it “did not take payment from Anthropic.”)

The headline numbers:

• Scores on AECI, Anthropic’s internal version of the Epoch Capabilities Index, have been growing at about 1.9 times its pre-October 2025 rate, with a 95% confidence interval of 1.5x to 2.8x. METR’s words: “Anthropic has made roughly two years of historical progress over the last year.”
• METR’s overall judgment is roughly 2x, but “we would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate.”
• On cause: “We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.” It is “confident that the acceleration was not due to an acceleration in the scaling of training compute.”
• On what slows it down: METR “did not see evidence for significant bottlenecks” in algorithmic progress, training compute, experiment compute or data over the next 12 months, while warning that its view of bottlenecks “changed several times.”
• The data includes three internal-only models that finished training between July and October 2026. The most recent, which METR calls Model 4, scores above 50% on every benchmark in the index it was tested on. METR says that inclusion is the main reason its figure is higher than the 1.7x reported in the Haiku 5.5 system card.

And the caveat that should travel with every one of those numbers: “Anthropic redacted a majority of the evidence we found most informative for these conclusions.” Anthropic described its redactions as “aimed at removing confidential information that could accelerate other AI developers.” To be fair to both sides, METR also says that “broadly the access provided surpassed our expectations,” and that staff-level access would have allowed a much deeper assessment still.

Two halves of a definition, and where METR’s numbers land

The current policy is Responsible Scaling Policy version 3.4, effective July 8, 2026 (the policy page lists every prior version). Its automated R&D threshold is met if Anthropic determines either 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)” or that there is “dramatic acceleration” of AI progress likely tied to automating AI research.

The second path is the one the filing points at, and it has two conditions. First, Anthropic must “observe or expect double the rate of progress in AI aggregate capabilities compared to both the rate we’d expect and the fastest rate of extended progress we’ve observed in the absence of significant AI contributions to AI R&D.” A footnote defines doubling as “as much progress in one year as one would see in two years at baseline,” and another defines “extended” as “over at least three model generations.” Second, it must be “plausible that this doubling is substantially attributable to the automation of research and/or engineering.”

Lay METR’s findings over that text and the picture sharpens:

• Attribution looks close to settled. An 80% AI share, with training compute ruled out, is the kind of evidence condition two asks for.
• The rate is sitting on the line. A 1.9x central estimate with an interval running from 1.5x to 2.8x straddles 2x. The filing’s “may imminently” phrasing is an honest description of a confidence interval, not hedging.
• The baseline is the live dispute. The policy measures against the fastest extended rate observed without AI help. METR flags that its numbers “could be overstated if the historical rate of growth was unusually low,” notes that Anthropic’s growth before late 2025 trailed the broader industry frontier, and shows that plotting all frontier models instead of only Anthropic’s gives roughly 1.4x instead of 1.9x. METR offers two readings of that gap: either Anthropic was subfrontier in 2025 and 1.4x is the truer number, or earlier models were more overfit to benchmarks and 1.9x stands. Which baseline Anthropic adopts could decide the designation.
• Version 3.4 already narrowed the trigger. The July changelog says Anthropic would not count the threshold as crossed if the overall pace of progress were constant or slowing, even if Anthropic estimated that pace was dramatically faster than it would be without advanced AI tools. It also admits the threshold “is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize.”

If you want the conceptual background on why labs treat this particular capability as the dangerous one, the site’s explainer on recursive self-improvement and the June piece on Anthropic’s co-founder calling for a pause cover it. What is new this week is that the abstraction now has a measured number and a confidence interval attached.

Crossing the line opens the books; the brake is wired to competitors

Here is where most coverage stops short. Under version 3.4, a model that crosses the automated R&D threshold becomes “highly capable” in the policy’s own vocabulary. The firm consequence written into the Risk Report section is this: Anthropic “will complete a full external review process” with at least one outside reviewer whenever a Risk Report covers highly capable models and is “significantly redacted,” defined as redactions that leave a reader unable to “meaningfully assess whether they agree with our conclusions.” Risk Reports themselves come every three to six months, plus an off-cycle analysis within 30 days of determining an internally deployed model significantly exceeds previously analyzed risk levels on automated R&D or high-stakes misalignment.

That is a transparency mechanism. METR’s own report shows why it matters: the first independent look at exactly this question was published with the most informative evidence redacted.

The delay mechanism sits elsewhere, in Appendix A, and it is conditional on the competitive landscape:

1. If Anthropic is in the lead, meaning it has or will “imminently” have a highly capable model and “clear evidence that no other competitor will soon develop such a model,” it will require a strong argument that catastrophic risk is contained and “will delay AI development and deployment as needed to achieve this, until and unless we no longer believe we have a significant lead.”
2. If competitors with highly capable models have strong safety measures, Anthropic will meet or exceed their overall risk reduction posture, delaying development and deployment until it does.
3. If a competitor ships a better mitigation, Anthropic will make “a significant effort” to match it, but “will not necessarily delay AI development and deployment.”

The introduction to the policy explains the logic plainly: a developer that pauses alone could leave “a world that is less safe,” because “the developers with the weakest protections would set the pace.” The policy keeps a general option open (“we would strongly consider pausing development and/or deployment” even where no commitment triggers), and the company’s mitigation plans, including an “eyes on everything” logging regime for internal AI development, are listed as plans rather than gates. Its roadmap goals are described as “not hard commitments but rather public goals.”

So the honest summary is that crossing the line guarantees more scrutiny and conditionally guarantees a delay. The condition is a judgment about OpenAI, Google DeepMind and everyone else. Given that OpenAI benched GPT-6.1 Astra over autonomy concerns and Google gated Gemini 4 Carbon to cyber defenders, “nobody else is close” is not an easy claim for Anthropic to make, which makes scenario one, the strongest delay commitment, the least likely to apply.

What changes for teams that buy Claude

The S-1 is unusually direct that safety calls are now a product variable. Future precautions “could include additional measures that affect whether, how, and when we develop or deploy new models or capabilities,” and Anthropic “cannot predict with precision when we will determine that such measures are necessary, what measures will be required, or how long they will remain in place.” It cites Mythos, released in April only as a gated preview through Project Glasswing, as the template, and separately warns that organizations “not selected for, or removed from, such programs” may see access decisions as arbitrary.

For an enterprise buyer, three practical readings follow:

• The best Anthropic model is not the one you can buy. METR’s data includes three internal models trained after July. Plan capability roadmaps on released models and treat the next release date as uncertain, not late.
• Gated launches are the new default, not the exception. If your use case needs a frontier capability on day one, the Mythos precedent says ask your account team now what vetted-access programs exist and what qualifies you, instead of finding out at launch.
• Read the next Risk Report like a supplier audit. The policy now requires it to mark where material was redacted. A report covering a “highly capable” model with heavy redactions has to carry an external reviewer’s verdict; that verdict is the closest thing a customer will get to an independent opinion on the vendor’s controls.

The same logic applies inward. If Anthropic, with every incentive to move fast, is writing human approval into its own R&D loop as models take over more of the work, the banks scoping which agent output enters model governance are not being timid. They are answering the “certain decisions” question for their own pipelines first.

The dates to watch are concrete. Anthropic’s last Risk Report was the August 2026 edition, so the next is due by February under the three-to-six-month cadence, sooner if an off-cycle analysis is triggered. METR has also published a separate review of that August report’s AI R&D section, which is worth reading alongside the next one. And the S-1 still lacks a price range; any further amendment that changes the “imminently” sentence, in either direction, will say more about the threshold than a press release would.