AI self-improvement is imminent? Anthropic warns in its prospectus

On November 2, Anthropic filed an amended S-1/A prospectus with the SEC, and its risk factors say AI self-improvement may be closer than it seemed. Based on the company’s internal estimates and its expectations for future models, Anthropic could reach, at any moment, the threshold that its own safety policy ties to AI self-improvement. TechCrunch was the first to point this out.

What exactly did Anthropic write?

The key passage says the company “may soon” meet the definition in its Responsible Scaling Policy (RSP) of “dramatically accelerated progress attributable to AI”. It immediately adds several caveats:

1. none of its current models has been determined to cross that threshold, and the company is still analyzing the data,
2. the assessment concerns future models, not the ones we use today,
3. Anthropic admits that the evidence of acceleration consists mostly of lagging indicators, so in the past it noticed jumps in pace only after the fact, and that could happen again,
4. the company promises further reports, as the RSP requires.

Elsewhere in the prospectus, Anthropic says it does not expect to be able to safely let Claude fully autonomously design and build future generations of models “without meaningful human oversight or understanding”. That is why it is considering changes to the RSP meant to keep humans in control of model development. What those changes are is not yet known.

What is AI self-improvement?

The term is “recursive self-improvement” (RSI): a situation in which AI does so much of the work on the next model that each generation speeds up the creation of the following one. The prospectus uses this term directly and calls building better AI with AI a “serious safety risk”, especially when it sharply speeds up the pace of progress.

In the RSP, this threshold is called the “AI R&D automation” threshold. According to the RSP update history, in version 3.1 from April Anthropic gave an example of the scale: compressing two years of AI progress from the 2018–2024 period into a single year, where what counts is the pace of AI capability progress as a whole, not the productivity of individual researchers. In July, in version 3.4, the company refined the threshold further.

How much is Claude already building Claude?

A lot. Figures from the same prospectus:

1. in September 2026, Claude wrote more than 90% of the code changes at Anthropic,
2. in Q3 2026, the average engineer contributed more than 45 times as much code per day as in 2024,
3. on internal engineering tasks without a clear specification, Claude’s success rate rose to about 90% in September, up about 50 percentage points in nine months (according to the company’s internal benchmarks),
4. the value of internal Claude usage at list prices grew more than 250-fold between January and September 2026, and in April it exceeded the company’s monthly personnel costs (excluding stock-based compensation) for the first time.

In other words, the tokens Anthropic consumes itself are worth more at list prices than all the salaries at the company. The prospectus says every version of Claude helps build the next one, and that compute devoted to internal use “continually accelerates” the pace of development.

Is this new in the AI world?

The topic itself is not. In June, Anthropic published the essay “When AI builds itself”, which said that in May more than 80% of the code merged at the company came from Claude, but that self-improvement had not happened yet (“We are not there yet”). It also proposed the option of slowing down or temporarily pausing model development, provided other labs did the same in a verifiable way.

What is new is where and how it was written. This time it is not an essay but a prospectus filed with the SEC, and the wording has shifted from “not yet” to “we may at any moment”. Interestingly, other labs are showing at the same time how models are starting to tinker with other models: yesterday I wrote about how GPT-6 Astra fixed Qwen’s weights, and earlier about Recursive Flint, which, according to the company, an automated research system helped build.

What should you watch for?

First, the context. This is a risk factor in a prospectus ahead of an IPO, and as TechCrunch notes, skeptics see the S-1 as partly a marketing document. Meanwhile on Hacker News, a lot of attention went to a piece by Matt Scherer of the Open Markets Institute urging the SEC to halt Anthropic’s IPO, among other reasons because of unaudited results for 2026. I covered the prospectus figures themselves in my post on Anthropic’s Q3 revenue.

Second, the RSP changes. The promise that humans will stay in control of model development sounds good, but there are no specifics yet. It is worth watching whether the new rules actually limit anything, or just describe what is already happening.

Third, the admission of a blind spot. The company itself writes that it may notice acceleration only after the fact. For us that means faster new models (which is of course welcome), but also the risk that safeguards will not keep up with capabilities.

FAQ

Is Claude already improving itself?

Not in the sense of full self-improvement. According to the prospectus, no current model has crossed the RSP threshold, although Claude already writes more than 90% of the code changes at Anthropic and helps build its next versions.

What is RSI in AI?

RSI (recursive self-improvement) is when a model speeds up the building of its successor, which in turn speeds up the building of the next one even more.

What is the Responsible Scaling Policy?

It is Anthropic’s policy that defines model capability thresholds and the safeguards required once they are crossed. One of those thresholds is AI R&D automation, which the prospectus refers to.

When will Anthropic cross the threshold?

The company does not say. The prospectus only states that it may happen “imminently” and that Anthropic will publish further reports under the RSP.