Derived from Lagging Indicators

In our piece last weekend on whether AI is starting to compound Anthropic's research, we quoted the company's amended prospectus, which said Anthropic "may imminently meet" a threshold in its safety policy. The relevant threshold concerns AI-assisted research doubling the pace at which AI capabilities improve. On November 10, Anthropic filed a second amendment saying it "made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models."

The system card for Claude Mythos 5.2, Anthropic's newest released model, is dated the day before and says Anthropic doesn't consider that model to have met the threshold, "though this assessment is uncertain." One document is about a single model and the other is about where the company expects its research to go, and both can be true.

Alongside the new determination, Anthropic still says, "We continue to assess the underlying data." It is also "evaluating the implications" of meeting the threshold and says further reports will follow.

Three versions of one paragraph

Before changing its threshold assessment, Anthropic revised its account of whether it could recognize acceleration in time. The original filing on October 26 warned that "we may not identify a significant acceleration in our rate of development until after it has occurred." On November 2, Anthropic added that it had "in the past not identified" such an acceleration until afterward. That acknowledgment remains in the November 10 version.

It also drops the line "We have not designated any of our current models as exceeding this threshold." We found no separate announcement of the determination in Anthropic's newsroom or policy page, whose latest dated entry was its August risk report. The Mythos 5.2 system card said Anthropic planned "to publish more soon on overall trends in acceleration."

Met by expectation

Anthropic's Responsible Scaling Policy sets out two ways to meet its automated-R&D threshold. One is a model that could stand in for all of Anthropic's research scientists and engineers at competitive cost. The Mythos 5.2 card says that model does not seem close to doing so.

The other concerns the pace of improvement across models. The policy defines a doubling as "as much progress in one year as one would see in two years at baseline," and notes that this "is not the same idea" as doubling researchers' productivity. It compares the pace with both what Anthropic would expect without significant AI contributions to research and the fastest extended progress it has observed without them. It must be plausible that automation of research or engineering accounts for a substantial part of the doubling, rather than factors such as more staff or compute. Continuing the trend must also seem likely to produce even greater acceleration.

The second path applies when Anthropic will "observe or expect" that doubling. The filing says its use of AI systems "has already been accelerating our research and development," while grounding the threshold determination in expectations for future models and its overall rate of progress. Allowing expectations matters when much of the evidence, as the filing puts it, "is derived from lagging indicators." By the time a benchmark trend shows faster progress, the research that produced those models has already taken place. Waiting for those results to establish a faster trend can leave a safety decision behind the work it is meant to govern.

Anthropic's own record shows how close the call has been. Its August risk report said leading indicators pointed to acceleration "by less than a factor of 2." It was fairly confident that the 2025 speedup came from factors other than AI, "though we also believe that our AI models have been a key factor in the faster trends continuing through the coverage date." METR, the nonprofit evaluator that reviewed that section with access to the unredacted report, thought the public version understated the acceleration. By METR's calculation, the private report's own figures imply a factor of 1.97, and METR considers it plausible that the threshold had been crossed by July 15. It adds that most of the relevant researchers at METR and at Anthropic lean toward the threshold not having been crossed then.

Those disagreements were about acceleration that had already shown up in Anthropic's data, which an outside evaluator could at least examine. A determination based on expectation depends on forecasts, and some of Anthropic's forecasts remain outside the public record. In its broader assessment, METR lists "forecasts of metrics for future models" among the material Anthropic's redactions kept it from communicating. Anthropic described those redactions as "aimed at removing confidential information that could accelerate other AI developers." Readers have the determination, but not enough of the forecasting evidence to assess its basis for themselves. The filing doesn't say which future models Anthropic has in mind, how much acceleration it expects from them, or when.

What Anthropic Plans to Do at the Threshold

As Anthropic uses Claude to accelerate its own work, the filing warns that the organization "may lose visibility across certain processes in a way that may become difficult for any individual or group of humans to understand and control." The policy's plan for an "eyes on everything" state addresses that risk by recording critical AI-development activity and using AI to look for security threats and concerning behavior by insiders, "humans as well as AI systems themselves."

The company also plans ambitious security research and systematic alignment assessments using interpretability and adversarial testing. It aims to make its internal testing of deployment safeguards better at finding jailbreaks than the combined efforts of participants in its established bug-bounty programs. Risk reports are to describe progress and findings, subject to external review. The policy separates these company plans from more demanding recommendations for the whole industry. Of those recommendations, Anthropic says, "we cannot commit to following them unilaterally."

Anthropic's Frontier Safety Roadmap puts dates on several of these goals and on related policy work. It called the capability plausible "as soon as early 2027" and said that "by that point, we expect to have accomplished most of the goals listed above." In October, the company reported that it had broadly achieved the first-phase goals of its security moonshot projects and completed its internal red-teaming goal ahead of schedule. The "eyes on everything" monitoring goal has a January 1 target. The related policy work has separate deadlines: December 15 for a proposal on embedded third-party evaluators, and January 15 for a framework on the pace of AI development or an explanation of why Anthropic decided against proposing one. None of those dates has passed, and the filing gives no timing for the models it expects.

The filing's concrete examples of restraint predate the determination or remain described as proposals. Anthropic writes that it recently delayed pre-training a new frontier base model "for more than a month" to strengthen protections against theft of its model weights, a disclosure that was already in the October filing. The November 2 amendment added that the company is considering updates to its policy "to require human review and approval of certain decisions in our research and development processes," and elsewhere that it is considering changes to "keep humans in control of frontier model development at Anthropic." The November 10 version still describes both as under consideration.

The same document shows the scale of Anthropic's own use of Claude. The value of Anthropic's internal Claude usage, priced at list rates, grew more than 250-fold from January to September 2026, and in April it passed the company's monthly personnel expense, excluding stock-based compensation. The filing ties that usage to each version of Claude helping to build the next. Anthropic also warns that precautionary measures "could slow our capability gains relative to competitors that do not adopt similar constraints."

Is the Policy Working?

In August, Anthropic said it was aiming to level up its risk mitigations "bearing in mind that we may cross this threshold in the coming year." Its November determination falls within that window, and its disclosed training delay shows that security concerns have already changed an internal development decision. A company acting on an expected risk can begin responding before a current model supplies decisive evidence; the absence of that evidence does not, by itself, make the determination premature.

The objection from the other side is that prospectus risk factors are written to anticipate everything that could go wrong, so a sentence there may say less about the state of Anthropic's research than the headlines that followed it. The filing's paragraph, though, went through three versions in fifteen days with each edit moving the same way, which reads as a revised conclusion.

Anthropic has decided that waiting for clearer evidence would mean waiting too long to recognize the risk. Its next risk report should explain what it did with that judgment: which protections it strengthened, which research it allowed to proceed, and whether anything had to wait. The purpose of calling the threshold on expectation is to change those decisions while there is still time, before the next generation of models produces another set of lagging indicators.