What Recursive Self-Improvement Means and How Close Anthropic Says It Is

Anthropic's revised securities filing, as reported by TechCrunch, says it may be near a model capable of recursive self-improvement, though no model has crossed its threshold. The label bundles distinct ideas, the measured evidence shows AI helping build AI in limited ways, and experts disagree about whether the loop would accelerate or stall.

Recursive self-improvement, usually shortened to RSI, describes AI systems that help make the next generation of AI better, which then helps make the generation after that better still. On Nov. 2, 2026, TechCrunch reported that Anthropic's revised S-1/A, a securities filing made ahead of a possible public offering, warns that the company is on the cusp of a model capable of RSI. According to the same report, Anthropic also said that none of its models has crossed its RSI threshold and that it promises to keep humans in the research and development process for the foreseeable future.

The topic matters because Anthropic, OpenAI and Google DeepMind have each built part of their safety policy around some version of this idea. The headline shorthand, "self-building AI," suggests a machine rewriting its own code. The labs' formal thresholds are more specific: they track how much AI speeds up AI research, or how much of that research it can automate. The distance between the slogan and the policy language is where most of the confusion, and most of the real debate, sits.

Four ideas hiding behind one label

The research behind this article separates four concepts that headlines tend to merge. Treating them as one makes it hard to judge how close anything is, because the evidence for each differs sharply.

1. AI-assisted R&D. AI tools help human researchers and engineers. This is happening now; Root-Nation, quoting the S-1, reports that Anthropic says it already uses AI in its research and development.
2. Full automation of AI research. AI performs the work of researchers end to end. This is the kind of capability the labs' formal thresholds describe.
3. Self-sustaining acceleration. AI speeds up AI progress without needing growth in outside inputs such as more chips or more people.
4. Intelligence explosion. Capabilities rise without limit in finite time, the theoretical endpoint of an argument made decades ago.

The last idea has the oldest pedigree. The statistician I.J. Good wrote in 1965 that an ultraintelligent machine could design even better machines, after which there would unquestionably be an intelligence explosion. A 2026 preprint titled "The Economics of Recursive Self-Improvement", which has not been peer reviewed, points out that these terms are often conflated with one another and with RSI itself. It defines self-sustaining acceleration and intelligence explosion as separate things and argues that an explosion need not follow if ideas become harder to find or if bottlenecks bind.

Per the research brief, the filing language and the lab thresholds mostly concern the first three ideas, not the fourth. That distinction shapes how the Anthropic disclosure should be read: a warning about AI-attributable acceleration is not a claim that a runaway process is under way.

What Anthropic said, and what it left unsaid

The filing as reported

The primary document, the SEC filing itself, was not available for this article, so every claim about it comes from news coverage. TechCrunch says the revised S-1/A, filed on a Monday, treats RSI as both a promise and a peril. It also says the filing points to Anthropic's Responsible Scaling Policy (RSP) as the framework for monitoring and reducing risk as the company approaches RSI, and that Anthropic is considering RSP updates to keep humans in control of frontier model development.

Gizmodo describes the new risk-factor text as a revision of an Oct. 26 filing and says RSI may arrive "imminently," a word choice that is Gizmodo's. Many other outlets that covered the story, including DEV Community, Techzine, Root-Nation and Free Press Kashmir, largely rewrote TechCrunch and therefore do not count as independent confirmation.

Reading "very close" carefully

The phrase "very close" rests on estimates, not on a crossing. According to secondary coverage, Anthropic's internal estimates about future models suggest it could soon meet the dramatic-acceleration trigger in its RSP. That is a forecast about models that do not yet exist, made by the company that would be affected by the trigger, and the language in the filing is hedged.

One reported line from the filing addresses the difficulty of knowing. Shiftdelete.net, a secondary outlet and the only one in the research to carry it, says Anthropic wrote that it has "in the past not identified, and may in the future not identify" a significant acceleration until after it occurred. The same outlet describes Anthropic as conceding that its evidence consists mostly of lagging indicators. If accurate, that means a threshold could be passed before anyone inside the company can confirm it.

Two other figures circulate and both are unverified. Shiftdelete.net reports that more than 90% of contributions to Anthropic's codebase in September 2026 were written by Claude. The SuperDataScience podcast puts the figure at more than 80% as of May 2026. Neither could be checked against the filing, and the share of code written is a measure of volume, not of research judgment or discovery.

The incentive question

TechCrunch itself raises the skeptic's reading: an S-1 is both a disclosure document and a marketing document, so a claim to be near a model that builds itself could serve as marketing. The claim is self-reported, made in a securities filing, by a company with a financial interest in being seen as at the frontier. Securities filings also carry legal pressure to disclose risks, which pulls the other way. The marketing caveat is a reason for caution, not a rebuttal, and the research brief rates its own confidence in the underlying Anthropic claim as moderate.

How the labs draw the line

Each of the three leading labs has written down a point at which AI-driven research becomes a safety matter. The definitions differ in wording and in how high they set the bar, which is one reason a statement like "close to RSI" means different things depending on whose policy is meant.

Anthropic

An earlier version of Anthropic's policy, as indexed by the Emerging Technology Observatory, defined the capability as the ability to fully automate the work of an entry-level, remote-only researcher at Anthropic, or to cause dramatic acceleration in the rate of effective scaling. According to the governance think tank GovAI, version 3.0 replaced the two AI R&D thresholds with a single one: a model capable of compressing two years of 2018–2024 AI progress into one year. Gizmodo describes the trigger as doubling the baseline rate of improvement, plausibly attributable to AI, but that wording was not confirmed in the policy text.

The current wording is unresolved. Anthropic's policy page says version 3.4, effective July 8, 2026, revises the automated R&D threshold "to better track the threat model of concern." The same page still refers to an AI R&D-4 threshold and says Claude Opus 4.6 does not cross it, adding that Anthropic will build an affirmative case about misalignment risks once models do. That conflicts with GovAI's account of version 3.0, and the discrepancy could not be settled without reading the full text of version 3.4. This article therefore does not state Anthropic's exact current threshold.

OpenAI

OpenAI's Preparedness Framework, version 2, published on April 15, 2025, tracks "AI Self-improvement" as one of three categories, alongside biological and chemical risk and cybersecurity. According to a 2026 preprint, "Measuring AI R&D Automation," the High level means an impact equivalent to giving every OpenAI researcher a highly performant mid-career research engineer assistant, measured against 2024 baselines. A METR compilation of safety policies excerpts the Critical level as a model "capable of recursively self improving (i.e., fully automated AI R&D)," though the full definition was truncated in the retrieved text.

The framework attaches duties to each level. Covered systems that reach High need safeguards before deployment, and systems that reach Critical need safeguards during development as well. In an August 2025 paper by OpenAI authors on worst-case risks of open-weight models, the company judged its gpt-oss model to clearly lack the capabilities required for a High classification.

Google DeepMind

Google DeepMind's Frontier Safety Framework includes "Machine Learning R&D" as a Critical Capability Level domain, according to the same 2026 preprint. It recommends high security measures for models that could significantly accelerate AI development. That description is secondhand, because the framework document itself was not retrieved, and no upper tier was available to compare.

Why a threshold, and why this one

The stated rationale in the lab policies and the filing, as reported, is that AI accelerating AI could outrun the safety work meant to keep pace with it. A second reason follows from Anthropic's own admission about lagging indicators: if such an acceleration is recognized only after it happens, a rule that waits for proof arrives late. Setting a written trigger in advance, tied to commitments such as safeguards before deployment, is meant to force a decision before the evidence is conclusive.

Whether the underlying risk is as large as the policies imply remains unresolved. The policies establish what the labs have committed to; they do not establish how likely or how fast the feared acceleration is.

What the measurements show so far

Longer tasks, faster

The most-cited trend line comes from METR, a nonprofit that evaluates AI systems. In a January 29, 2026 update, METR reported that the length of software task that frontier models complete with 50% success has been growing exponentially. Its original 2019–2025 dataset showed a doubling about every 7 months, and its revised hybrid version, called TH1.1, gives a nearly identical 196-day doubling time over the full period. Measured from 2024 onward, the doubling was faster: 109 days under the original task set and 89 days under TH1.1.

Brackets are confidence intervals where METR reports them. A tracker built around the AI 2027 forecast, citing the AI Futures Project's first-quarter 2026 update, reports a 4–4.5 month doubling in forecasters' revised models. That tracker is an advocacy source and is used here with caution.

The metric has limits that bear directly on the RSI question. Time horizon reflects how long human experts take on tasks that an AI can finish; it is not a measure of how fast AI does research, and it covers software tasks only. METR says the trend is somewhat sensitive to task composition, and a critical blog post argues the fast recent estimates are noisy because few tasks sit above 8 hours. METR's rise in long tasks from 14 to 31 eases that concern without removing it. The brief treats the trend as moderately to highly reliable, and any precise rate or extrapolation as low confidence.

A real loop, still a small one

The clearest documented case of AI improving the systems that build AI is Google DeepMind's AlphaEvolve, announced in May 2025. According to DeepMind's blog post, the coding agent sped up a vital kernel in Gemini's architecture by 23%, which cut Gemini's training time by 1%. It also achieved a speedup of up to 32.5% on the FlashAttention kernel.

VentureBeat quotes DeepMind researcher Alexander Novikov putting the kernel gain at 1% in end-to-end savings on the entire Gemini training run. This is a self-referential loop that is real but small, and humans directed it. A 2026 preprint on recursive self-improvement names the automatic scorer, or evaluator, that such systems depend on as a recurring bottleneck: AlphaEvolve works where speed and correctness can be checked automatically, which is a narrower setting than open-ended research.

Taken together, the available evidence shows AI helping to build AI, in kernels and in code, rather than autonomously building a successor. The brief rates that conclusion as moderate in confidence, partly because internal lab data are not public. No randomized trial of AI uplift for frontier AI researchers turned up in the research, and such data would show most directly whether acceleration is real.

Why skeptics expect friction

The case against a runaway loop rests on three kinds of constraint, and the sources for each are preprints and economic models that have not been peer reviewed. They depend on assumptions about quantities nobody has measured well, so they are arguments to weigh, not results.

• Compute. Whitfill and Wu's 2025 preprint, "Will Compute Bottlenecks Prevent an Intelligence Explosion?", estimates how substitutable research compute and cognitive labor are at frontier labs. If the two are gross complements, more AI labor will run into the compute limit and recursive improvement will fizzle. The retrieved text did not state their estimate, and the paper was not read in full.
• Diminishing returns. The 2026 economics preprint says an explosion may not follow if ideas become harder to find. The brief also notes that a paper cited by "Measuring AI R&D Automation," by Erdil and Barnett (2025), argues that automating AI R&D alone likely will not accelerate AI progress; that argument was not retrieved directly.
• Novelty. A separate preprint argues that routine machine-learning engineering can be parallelized, while generating novel insight is fundamentally serial.
• Compute-slowdown modeling. A 2025 forecasting preprint finds that, in its model, the time-horizon growth rate is causally proportional to the compute growth rate. It excludes a software-only singularity by assumption, and its authors flag that assumption as contestable.

Counterevidence also comes from the labs themselves. OpenAI's August 2025 judgment that gpt-oss fell clearly short of its High level shows that at least one lab found a model well below its own line. METR's own limits note that benchmarks differ from real research, so a model can be strong at coding and still fail to run research end to end.

The bottleneck arguments do not show the loop cannot happen. They show that the outcome hinges on how substitutable compute and labor are, how quickly returns diminish and how well AI systems can evaluate their own work. According to the brief, none of those is settled empirically, so whether feedback loops turn explosive or stall is unresolved.

What it means and what to watch

The most defensible reading of the Anthropic disclosure is modest and still significant. A leading lab has put in a securities filing that its own estimates point toward a threshold it defined for itself, that it has not crossed it, and that it may recognize a crossing only in hindsight. That is not evidence of a machine rewriting itself, but it is a notable statement about how the people closest to the work see the trajectory.

What would sharpen the picture is the kind of evidence the labs have mostly not published: independent audits, evaluations by outside groups such as METR or the UK AI Security Institute, and the labs' own studies of how much AI actually speeds up their researchers. Also worth watching are the text of any RSP revisions meant to keep humans in control of development, and whether other labs disclose comparable estimates. Until then, the question that matters is less whether a model builds itself than whether anyone can tell, in time, that AI has started to speed up its own progress.

Frequently asked questions

What is recursive self-improvement in AI?

It refers to AI systems that improve the process of building AI, so that each generation helps produce a better next one. In lab policies it usually means AI that accelerates or automates AI research to a defined degree, rather than a model literally rewriting its own code. Researchers distinguish it from self-sustaining acceleration and from an intelligence explosion, which are separate and stronger ideas.

Has Anthropic built a self-improving AI?

According to TechCrunch's report on Anthropic's revised S-1/A, no. Anthropic said none of its models has crossed its RSI threshold, while warning that it is on the cusp of a model capable of RSI. The "very close" language rests on internal estimates about future models, per secondary coverage.

How do OpenAI and Google DeepMind define the threshold?

OpenAI's Preparedness Framework version 2 tracks AI self-improvement, with a High level equivalent to giving every researcher a mid-career research-engineer assistant and a Critical level of fully automated AI R&D. Google DeepMind's Frontier Safety Framework lists machine-learning R&D as a Critical Capability Level domain, according to a secondary description. Anthropic's threshold is based on AI-attributable acceleration of progress, though its exact current wording could not be confirmed.

Is there evidence that AI is already speeding up AI development?

There is limited, documented evidence. Google DeepMind reported that AlphaEvolve sped up a Gemini kernel by 23%, which cut training time by 1%. Claims that Claude writes more than 80% to 90% of Anthropic's code are unverified, and code volume is not the same as research progress.

Why do skeptics doubt that self-improving AI would explode in capability?

Preprints argue that compute may limit the loop, that ideas become harder to find, and that novel insight is hard to parallelize. Whitfill and Wu examine whether research compute and cognitive labor are complements, which would cause the loop to fizzle. These arguments rest on assumptions that have not been settled by evidence.

Sources

1. Anthropic warns that it's very close to RSI capability, aka self-building AI , TechCrunch (2026-11-02)
2. Anthropic Claims to Be Near Recursive Self-Improvement , Gizmodo
3. Anthropic Says Recursive Self-Improvement May Be Imminent , Shiftdelete.net
4. Anthropic Discusses Recursive Self-Improvement , Root-Nation.com
5. Anthropic's Responsible Scaling Policy , Anthropic
6. Anthropic Responsible Scaling Policy (earlier version, indexed by AGORA) , Emerging Technology Observatory (AGORA)
7. Anthropic's RSP v3.0: How it Works, What's Changed, and Some Reflections , GovAI
8. Our updated Preparedness Framework , OpenAI (2025-04-15)
9. Common Elements of Frontier AI Safety Policies , METR
10. Time Horizon 1.1 , METR (2026-01-29)
11. METR time horizon doubles every 4 months , AI 2027 Tracker
12. AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms , Google DeepMind
13. Meet AlphaEvolve, the Google AI that writes its own code—and just saved millions in computing costs , VentureBeat
14. Will Compute Bottlenecks Prevent an Intelligence Explosion? , arXiv (Whitfill & Wu)
15. The Economics of Recursive Self-Improvement , arXiv
16. Measuring AI R&D Automation , arXiv
17. Forecasting AI Time Horizon Under Compute Slowdowns , arXiv
18. The Novelty Bottleneck: A Framework for Understanding Human Effort Scaling in AI-Assisted Work , arXiv
19. Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops , arXiv
20. Estimating Worst-Case Frontier Risks of Open-Weight LLMs , arXiv (OpenAI authors) (2025-08-01)
21. SDS 1004: Recursive Self-Improvement , SuperDataScience

This article was drafted with the help of AI and reviewed by Jaroha editors before publication.