What did Anthropic's amended filing say?
Anthropic amended its public listing paperwork on 2 November 2026. The revised Form S-1/A, filed with the US Securities and Exchange Commission, adds language to the risk factors section about how AI is changing the company's own research. TechCrunch reported the change the same day.
The key sentence concerns Anthropic's Responsible Scaling Policy, or RSP. That is the voluntary framework mapping model capabilities to safety requirements. The filing says internal estimates and expectations about future models, rather than current ones, suggest a threshold may be close. It states that Anthropic "may imminently meet" its RSP definition of "dramatically accelerated progress attributable to AI."
Anthropic also said no existing model has crossed that line. "We have not designated any of our current models as exceeding this threshold," the filing says, adding that the company continues to assess the underlying data. The passage gives no date for when a future model might reach it.
The phrase does not appear in the original S-1 filed on 26 October. A full-text search of EDGAR, the SEC's filing system, returns the sentence only in the 2 November amendment. That makes the disclosure a deliberate addition rather than recycled boilerplate.
The passage sits inside a risk document the Guardian described as 92 pages long. Much of it covers business and safety risks that any company preparing to list would disclose. The new passage stands out because it describes a safety threshold in forward-looking terms.
Anthropic uses a phrase for the underlying trend: "closing the loop." The filing describes a cycle in which each version of Claude helps design, train, test and optimize the next. The company calls Claude one of the most powerful drivers of its own progress, not a side experiment.
What is recursive self-improvement?
Recursive self-improvement, usually shortened to RSI, describes a loop in which an AI system helps build the next version of itself. Each cycle is meant to make the following one faster or cheaper. Taken to its limit, the loop would let models design and train their successors with little human input.
Anthropic's research arm puts the idea plainly. It says AI systems could become "capable of full recursive self-improvement, and begin building their successors." The same report stresses that this is a direction of travel, not a switch that flips. Today's systems can run experiments and write code. The remaining gap, Anthropic says, is judgment about which goals to pursue.
The concept has a long history in AI writing. Nick Bostrom's 2014 book Superintelligence treated a self-improving system as a central route to a rapid capability jump. What is new is the setting. A private company is now describing the idea in a document written for public investors.
The term is also contested. When RSI became popular earlier this year, many people treated it as a buzzword or a goal too vague to define or prove. Anthropic has given the idea a policy definition and a measurement test, and the filing now adds a public disclosure. That does not settle the scientific question, but it makes the claim checkable in principle.
Anthropic's chief executive has already made a similar argument in public. In a September 2026 essay, Dario Amodei wrote that AI has been advancing "drastically faster" since the summer, driven mainly by AI's ability to build the next generation. He said RSI is "starting to happen across the industry, including at Anthropic." He also called for the industry to slow the pace of capability improvements.
That public position matters for reading the filing. The amended language is consistent with warnings Anthropic has issued for months, not a one-off phrase inserted by lawyers.
How close is Anthropic to its own RSI threshold?
The RSP defines the bar in concrete terms. Its threshold asks whether Anthropic observes or expects to "double the rate of progress" in aggregate AI capabilities. The policy explains that as much progress in one year as two years at baseline. It adds that this is not the same as doubling researchers' productivity.
The threshold has two routes. One is full substitution for Anthropic's research scientists and engineers at competitive cost, defined as within a factor of five. The other is dramatic acceleration of AI progress. The filing refers to the second.
The framework has changed often. Anthropic published the first RSP in September 2023 and rewrote it in February 2026. It then issued revisions in April, May and July 2026. Each update adjusted thresholds, reporting rules or internal review. The frequency shows a company still deciding how to measure the risk it has promised to manage.
The latest version, 3.4, took effect on 8 July 2026. It says the threshold "is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize." The same update says Anthropic would not count a constant or slowing overall rate of progress, even if AI tools were doing more of the work.
That caveat matters. The filing's claim rests on internal estimates and expectations, not a measured breach. Anthropic concedes that its evidence is largely backward-looking. It says it may not identify a significant acceleration "until after it has occurred," because the signals arrive late.
The company also left itself room to act before any model crosses the line. The filing says Anthropic is evaluating what meeting the threshold would imply and will publish further reports under its policy. Separately, it says it is considering changes to the RSP to keep humans in control of frontier model development.
Crossing the threshold would not automatically stop work. The RSP pairs capability levels with additional safeguards, monitoring and reporting. Anthropic has also said it may pause development or deployment even when its formal commitments are not triggered.
What evidence does Anthropic cite for acceleration?
The filing's language is short. The supporting case is longer, and most of it comes from Anthropic's own systems and its research arm.
Anthropic also cites Project Glasswing, a limited release of Mythos. In its first weeks, the company says the model found more than ten thousand high- and critical-severity software vulnerabilities. That claim concerns security capability, not RSI, but it illustrates the pace argument.
The research arm adds two more signals. On the most open-ended coding tasks, Claude's success rate reached 76% in May 2026, up 50 percentage points in six months. In a March 2026 poll of 130 employees, the median respondent estimated producing about four times as much output with Mythos Preview. Anthropic says the true uplift was probably lower.
- •More than 80% of code merged into Anthropic's codebase was written by Claude as of May 2026. Before Claude Code's research preview in February 2025, the figure was in the low single digits.
- •In the second quarter of 2026, the typical engineer merged eight times as many lines of code per day as in 2024. Anthropic calls lines of code an imperfect measure and says the figure probably overstates the productivity gain.
- •The length of tasks Claude can complete has doubled roughly every four months, compared with every seven months earlier. Anthropic says Opus 3 handled four-minute tasks in March 2024. Sonnet 3.7 handled 90-minute tasks a year later, and Opus 4.6 handled 12-hour tasks a year after that.
- •METR, an independent evaluator, found that a Claude model called Mythos Preview could work for at least 16 hours. That was the upper end of what its tests could measure without new tasks.
- •In an April 2026 experiment, Claude-powered agents recovered 97% of the gap between a weak and a strong model on an AI safety problem. Two human researchers recovered about 23% in a week. The agents used 800 cumulative hours and about $18,000 of compute.
- •An automated Claude reviewer would have caught roughly one third of the bugs behind past incidents on claude.ai, according to a retrospective analysis.
Why does the disclosure matter for investors and regulators?
The filing also offers a financial measure of internal adoption. It says the value of Anthropic's own Claude usage at list prices grew more than 250 times between January and September 2026. In April 2026, that monthly value first exceeded monthly personnel expense, excluding stock-based compensation.
That figure is a proxy for how deeply the company relies on its own tools. It does not show that Claude can improve Claude without people. It does show that the tools are central to daily work, which is the premise behind the RSI argument.
The evidence still has limits, and Anthropic says so. Lines of code can rise while quality falls. A model judge can reward a plausible answer. A research experiment can succeed in a narrow setting and fail at production scale. The filing's conclusion depends on many such signals pointing the same way.
Risk factors carry legal weight. They are where a company describes what could go wrong, partly to limit legal exposure. Saying a safety threshold may be near is therefore a meaningful statement, even when it is hedged.
For investors, the timing is awkward. They are being asked to price a company that says its own tools are accelerating its research, while also warning that oversight may lag. The filing frames the risk in business terms. Safety incidents, regulatory intervention and reputational harm could all hurt results.
For regulators, the passage shows the limits of voluntary commitments. Anthropic's RSP is its own policy, not law. The company has separately proposed an Advanced AI Framework that would give governments authority to block or deter dangerous deployments. That proposal would apply to models trained with more than 10^25 floating-point operations, from companies above certain revenue or research spending thresholds.
For workers and researchers, the disclosure adds weight to a debate already under way inside the company. Wired reported that a researcher resigned from Anthropic in September, warning that advances including RSI could lead to human extinction within a few years. Anthropic has not endorsed that timeline publicly, but its filing acknowledges the underlying concern.
The disclosure also changes how competitors may be judged. If Anthropic puts a definition on the record, other labs face pressure to say whether they have an equivalent test and whether they have passed it.
What the filing does not show
The filing does not show that Anthropic has built a self-improving system. It says the opposite: no current model has been designated as crossing the threshold. The signal comes from forecasts about models that do not yet exist.
Most of the supporting evidence is self-reported. The code, task-length and experiment figures come from Anthropic's own systems and its own research arm. Anthropic discloses some limits, including that lines of code overstate productivity and that its April experiment did not transfer cleanly to production-scale models.
Some of the research results rest on model judges. In one comparison, a separate Claude model decided whether the AI or the human had suggested the better next step. Opus 4.5 beat the human choice 51% of the time in November 2025; Mythos Preview reached 64% in April 2026. The sample was chosen from moments when the human's choice had room for improvement, so it is not a like-for-like test of AI and human judgment.
The filing is also a sales document. TechCrunch noted that skeptics may read claims of near-RSI capability as positioning as much as disclosure. That does not make the claim false, but it argues for waiting for independent evaluation.
Finally, the threshold itself is a work in progress. Anthropic's own policy says it has been difficult to operationalize. A definition that shifts with each revision is a weak basis for saying a line has been crossed.
What to watch next
The clearest test is Anthropic's next Risk Report. The most recent public edition was published in August 2026. The filing says further reports will follow as the company evaluates the threshold. Those reports should show whether the internal estimates firm up or fade.
Watch the RSP itself. Anthropic says it is considering updates to keep humans in control of frontier model development. A revised policy could add new reporting duties, new evaluation requirements or explicit pause conditions. The details will matter more than the headline.
Watch the listing process. An IPO forces the company to keep disclosing material risks to public shareholders. If the threshold is crossed before or after listing, Anthropic will have to explain what changed and what safeguards followed.
Watch regulators. Anthropic's proposed framework would give governments authority to block or deter deployment of the largest, most capable models. Whether lawmakers adopt anything like it will shape how much the RSP matters outside the company.
Watch competitors. RSI is not an Anthropic-only question. If other labs publish similar thresholds, the industry will have a common yardstick. If they do not, Anthropic's disclosure will remain a single company's account of its own progress.