Recursive self-improvement: Anthropic says it has met the threshold

In its prospectus dated November 10, Anthropic wrote that it has determined it met its threshold for the automation of AI research and development. This is the threshold from the Responsible Scaling Policy (RSP) which, according to that policy, is meant to capture the onset of “dramatic recursive self-improvement”. As recently as November 2, the same passage of the prospectus said only that the company might reach the threshold soon. Yahoo Finance spotted the change.

What exactly changed in the prospectus?

One sentence in the risk factors. In the new version, the company’s determination rests on what it expects from future models and on its overall rate of progress, not on what it has seen in current models (“we 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”).

That distinction matters. The RSP threshold covers a situation in which the company would “observe or expect” the acceleration, so a forecast is enough to meet it. Anthropic is therefore not claiming that Fable 5.2 or Mythos 5.2 are already improving themselves. It is claiming that, given what it sees internally, it expects that pace.

Further on, according to the prospectus, the company is still analyzing the data, assessing the implications of meeting the threshold, and will publish further reports. I covered the previous version of this passage in my post on AI self-improvement in Anthropic’s prospectus.

Recursive self-improvement according to the RSP

In RSP version 3.4, in force since July 8, the threshold is met in either of two cases:

1. models are able to fully replace all of the company’s researchers and research engineers at competitive cost (within a factor of five),
2. there is a dramatic acceleration of AI progress, meaning the company observes or expects a doubling of the rate of AI capability growth, and it is likely that the doubling is in large part due to the automation of research and engineering.

The wording in the prospectus (“observe or expect”) comes from the second case. So this is about pace, not about replacing people. In the change log attached to the same RSP version, Anthropic also admits that the threshold is meant to capture the onset of recursive self-improvement and that it is hard to translate into concrete measurements.

What does METR say?

On November 5, five days before the new version of the prospectus, METR published two independent assessments. In the first, covering October 2026, it estimates that since October 2025 the company’s internal models have been improving 1.9 times faster than before (95% confidence interval: 1.5–2.8). The authors’ guess is that about 80% of that acceleration is due to AI, and they are confident it does not come from faster scaling of training compute.

The second is a review of the August risk report, in which Anthropic argued that it had not yet crossed the threshold. METR found that argument “overstated” and wrote that the doubling threshold may have been crossed as early as July 15, and that the public version of the report understates the scale of the acceleration.

Take these numbers with some distance: METR notes that in the public version Anthropic redacted most of the evidence the organization found most convincing, and that it had no direct access to unreleased models. For comparison, at the other leading labs METR sees a smaller acceleration, about 1.4 times.

What should you watch for?

First, the lack of specifics. The prospectus does not name the safeguards that meeting the threshold triggers. It lists only what the company has done so far, namely delaying releases and restricting access to models (as with Mythos in Project Glasswing), and warns that such steps may become more frequent and more costly. On the RSP updates page, the latest entry is from August 14, so as of today, November 12, we know about this decision only from a document written for investors.

Second, the calendar. According to the IPO page, pricing is expected on November 18, with trading to start a day later. A risk factor in a prospectus is also legal protection for the company, which is worth keeping in mind when you read it a week before the listing (I wrote about the range in my post on Anthropic’s stock price).

Third, the sentence about losing visibility. According to the prospectus, the more Anthropic relies on Claude, the more the organization may lose insight into some of its processes, in ways that are hard for any person or group of people to understand and control. For us that means faster new models (Fable 5.2 came out three days ago), but also the risk that the first response to the threshold will be slower releases and narrower access.

FAQ

Is Claude already improving itself?

Not fully autonomously. According to the prospectus, Claude wrote more than 90% of the contributions to Anthropic’s code in September 2026, but the company does not expect to be able to safely let it design and build the next models on its own without meaningful human oversight.

Have current models crossed the threshold?

The prospectus does not say that. The determination rests on expectations for future models and the overall rate of progress, not on observations of current models.

What changes now that the threshold has been met?

Nobody knows yet. Anthropic writes that it is assessing the implications, promises further reports, and is considering changes to the RSP meant to keep humans in control of model development.