Deep Dive — Anthropic says recursive self-improvement may be near

The most important sentence in Anthropic's amended IPO filing is not a number. It is a forecast, buried in the risk factors, that the company's own next models may cross the line it drew for AI that can meaningfully speed up its own development.

"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,'" the filing states. It adds, in the same paragraph, that no current model has been designated as exceeding the threshold, and that the company is still assessing the underlying data — per Amendment No. 1 to Anthropic's Form S-1, filed with the SEC on November 2.

What a reader would type to find this: "Anthropic recursive self-improvement", "Anthropic RSI", "Anthropic IPO filing". This is the first time a frontier lab has put that forecast in a document with underwriter liability attached.

What the filing actually says

The language is careful in a way that makes it more interesting, not less. Anthropic's own models are already doing research work inside the company, and the filing concedes that the feedback loop has a detection problem: increasing use of AI to accelerate its own R&D "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."

Then the admission that matters most: "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."

Read that twice. The company is not saying it will notice the moment its models start building better models. It is saying it has already missed a jump before, and expects the next one could also be visible only in the rear-view mirror. The forecast about future models is the reason it is saying so now.

The remedy it floats is procedural rather than technical. Anthropic says it is considering requiring human review and approval of "certain" research tasks currently handled automatically, and is weighing an update to its Responsible Scaling Policy "to keep humans in control of frontier model development." It also states plainly that it does not expect to be able to safely let Claude "fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible."

That last clause is doing a lot of work. It is a public commitment that autonomy in frontier R&D is a thing Anthropic will not do even if it can — written into a prospectus, where a later reversal is a disclosure event rather than a blog post.

What the threshold actually is

"Dramatically accelerated progress" is not a vibe; it has a definition inside Anthropic's Responsible Scaling Policy, and the definition is narrower than the doom framing suggests. The policy's AI R&D capability threshold can be met in one of two ways: models that could substitute for the company's research scientists and engineers at comparable cost, or a doubling in the rate of progress of overall AI capability that is plausibly caused by automated research rather than by hiring more people or buying more compute, measured against the fastest stretch of progress seen without significant AI help across at least three model generations — under the policy's own wording.

The doubling is the part to hold onto. It means as much capability progress in one year as the company would otherwise have gotten in two. If Anthropic's baseline was a ninefold effective scale-up in a year, the threshold is crossed when it expects something closer to eighty-one-fold — the policy's own illustrative arithmetic.

On that reading, "may imminently meet" is a claim about the slope of the curve, not about a machine waking up. It says the company's internal projections for models it has not shipped yet now point at a pace it has never seen. That is a stronger claim than the September rumor mill, and a weaker one than the headlines about self-building AI. It is also the first version of the claim that comes with a legal signature.

The context is that Anthropic's chief executive already said something similar in public. Dario Amodei wrote in September that AI had been advancing "drastically faster" since the summer, driven primarily by AI's growing ability to build the next generation — the essay that set off the round of RSI speculation we covered in Google won't confirm the RSI rumor. The filing turns the essay into a risk factor. That is what a company does when the engineers keep saying the same thing in meetings and the lawyers decide it has to be disclosed.

The other half of the filing is the same argument

The financial update published alongside that language is the reason to take the safety disclosure seriously rather than file it under marketing.

Anthropic booked $17 billion in revenue in the quarter ended September 30, roughly twelve times the same quarter a year earlier, according to SiliconANGLE's read of the filing. The growth rate, though, halved: revenue rose 47 percent quarter over quarter, down from 144 percent in the prior quarter. Net loss came in at $23.3 billion against $12.1 billion a year earlier, with operating expenses up 292 percent to $17.5 billion, per Bloomberg's read of the same filing.

Two charges explain most of the swing from the profit Anthropic reported in the second quarter — the first profitable quarter by its adjusted measure, when operating income was $974 million. In the third quarter the company took a $5 billion charge for buying computing hardware at a premium because demand arrived faster than its 12-to-18-month capacity planning expected, and a $2 billion charge tied to its non-cash employee donation matching program. That left a $536 million operating loss. Adjusted operating income, which strips out stock compensation and charitable contributions, was still $2.5 billion, at a 15 percent margin, down from 19 percent.

Then the balance-sheet numbers, which are the ones that decide whether any of this works. Anthropic listed $895.4 billion of outstanding infrastructure commitments as of September 30. Leases with Google and Broadcom account for nearly a third of that. Purchase commitments jumped from $55 billion at the end of 2025 to about $314 billion. Five weeks ago, the leaked draft of this same prospectus put the decade-long infrastructure promise at $518 billion — the number we worked through in Anthropic promised $518 billion. The public filing is 73 percent larger.

The offering itself is being pitched at a valuation of $1.5 trillion to $2 trillion, Reuters reported, with Bloomberg reporting the company aims to raise at least the $86.2 billion SpaceX pulled in earlier this year.

Put the two halves side by side and they are one argument. Annualized, $17 billion a quarter is $68 billion of revenue against $895.4 billion of commitments — about thirteen years of current revenue, most of it contracted whether or not customers show up. That arithmetic only closes if revenue keeps compounding at a rate no enterprise software company has sustained. The single mechanism that could produce that kind of step change is a faster model-development loop. Which is exactly what the risk factors say may be about to happen, and exactly what they say the company may fail to notice in time.

A shareholder buying this listing is therefore underwriting two bets that look opposite and are not: that Anthropic can accelerate, and that it can govern an acceleration it admits it may not detect.

Where the skeptics are right

Three objections deserve more than a shrug.

The first is that Anthropic wrote the definition, measures itself against it, and publishes the verdict. No auditor signs off on whether a model has crossed the AI R&D line. A threshold a company sets for itself is a useful internal control and a weak external commitment, and the filing's hedge — "future models, rather than our current models" — leaves the company room to say for years that the crossing is one release away.

The second is that an S-1 is a sales document with a legal appendix, as TechCrunch noted in its reading of the same passage. A lab that tells public investors it is close to a technology of unbounded upside, and also warns that the same technology is a considerable safety risk, has hedged in both directions at once. That is what the genre rewards. It is not evidence that the engineers are wrong.

The third is the top line. If recursive self-improvement were already underway, the first place it should show up is pricing power and revenue growth. Instead quarterly growth fell from 144 percent to 47 percent, and the company burned a $5 billion premium buying hardware it had not planned for. The $5 billion charge is acceleration of a different kind — the pace at which capacity has to be bought, not the pace at which capability arrives — and it is a cost, not a moat.

There is also a competitive fact the filing sits next to. Alphabet's newest model, Gemini 4 Carbon, outperforms Claude Opus 5.5 across multiple benchmarks, per SiliconANGLE. A company that is about to meet its own definition of dramatic acceleration is, on the public evidence, not obviously ahead of the lab with the most compute.

The case for believing it anyway

The strongest argument for taking the disclosure seriously is that it is an admission against interest. A prospectus is the one document where saying "we may not notice our own acceleration" costs something: it invites underwriter questions, regulator attention and a shareholder suit if the claim later looks like puffery. Nobody puts that sentence in a filing to look impressive.

The second is that it is falsifiable on a short clock. Anthropic says it will release further reports as it works through the implications of meeting the AI R&D threshold, and it is weighing an RSP update before the offering prices. If the update lands, it will say what human approval actually covers — evaluations, training runs, or the research decisions that choose what to build next. If it does not land before the roadshow, the disclosure was a risk factor and nothing more.

The third is that this stops being one company's story the moment a second lab files. OpenAI is expected to list. If its prospectus carries similar language about automated R&D and detection lag, RSI stops being a claim in a blog post and becomes a disclosed industry condition. If it does not, Anthropic's paragraph reads as differentiation aimed at investors who want both the upside and the conscience.

What to watch

• The RSP update, before pricing. The filing promises human review of "certain" research tasks. The word certain is where the argument will be. Watch whether it names training runs and evaluation design, or only deployment decisions.
• The first AI R&D threshold report. Anthropic committed to publishing its assessment. A report that describes the measurements is a control; a report that describes the conclusion is a press release.
• OpenAI's S-1 language. One lab disclosing a detection lag is a risk factor. Two labs disclosing it is the market pricing a new class of risk.
• Fourth-quarter growth. A second quarter of deceleration makes the acceleration forecast harder to sell. A re-acceleration, even a modest one, is the number that would make the filing's safety paragraph look prophetic rather than promotional.
• The $895.4 billion, and the Google clause. The September draft forced Anthropic to pay Google the difference if it underspent. If that clause survived into the public filing, the commitments are a floor, not a budget.

If a company tells you it may not notice its own models getting dramatically faster, should regulators take that at face value — or treat it as the most revealing sentence in the filing? Write to us via our contact page.