On November 2, 2026, Beijing time, artificial intelligence company Anthropic disclosed in its S-1/A filing with the U.S. Securities and Exchange Commission that its internal estimates for future models may soon reach the significant acceleration threshold defined by the company's policies. The document confines the scope of judgment to expected future models and explicitly states that the current model has not yet been determined to exceed this standard. The company is still evaluating underlying data, and existing disclosures cannot prove that recursive self-improvement has been achieved. This content appears in the registration document's risk factors section, alongside the company's description of its internal research and development processes.
Future Models May Approach Acceleration Threshold but Current Model Not Yet Exceeding It
Anthropic refers to the participation of consecutive version models in the development of next-generation models as 'closing the loop' in the document. The company states that Claude has been used for experiments, training data generation and processing, automated evaluation, safety audits, and code writing and review. The filing does not state that the current process has achieved complete autonomous design and construction of subsequent models, and explicitly notes that if such capability becomes technically possible, the company does not expect to be able to safely use it without substantial human supervision or understanding.
The risk factors section of the filing states that AI-assisted research and development may cause model capability advancement to outpace the development, validation, and implementation of safety, alignment, and security measures. Anthropic indicates that much of the evidence regarding this acceleration speed comes from lagging indicators, and the company lacks complete predictive and measurement capabilities for acceleration brought by recent or unreleased models. The company also acknowledges that it has previously identified significant acceleration only after it had already occurred, and may not be able to identify it in a timely manner in the future.
Description and Risk Boundaries of AI-Driven R&D Processes
Regarding policy thresholds, the document states that judgments are based on internal estimates and expectations of future models, rather than existing models. Anthropic indicates that future models may soon reach the 'significant accelerated progress brought by AI' threshold defined by its Responsible Scaling Policy, but current models have not been designated as exceeding this threshold. The company is evaluating the impact of reaching AI development capability thresholds and plans to continue publishing reports according to this policy, with specific underlying data not being disclosed. The company has also not announced when the evaluation will be completed or provided a list of future models used for this determination.
Anthropic simultaneously disclosed that its reliance on its own AI systems has expanded to production code, research, and the construction of subsequent model versions. The document writes the resulting risks into the filing: systematic defects, biases, or errors in current models may be propagated to next-generation models and products, and accumulate across consecutive versions, becoming difficult to detect or repair. This is the risk scenario listed by the company, with public materials not indicating that related propagation has already caused specific incidents.
Company Acknowledgment of Lagging Indicators and Identification Delays
At the research and development process level, the filing documents state that Claude participated in designing, training, testing, and optimizing next-generation AI systems, and was used for running experiments, generating and processing training data, conducting evaluations and safety audits, and writing and reviewing code. These descriptions come from Anthropic's self-reporting of internal working methods, and the documents do not provide external audits broken down by task, nor do they equate the aforementioned participation methods with the model being capable of completing the entire R&D cycle autonomously without human intervention.
The documents also list three sets of internal operational data. Anthropic self-reports that over 90% of the company's codebase contributions in September 2026 were written by Claude; for the three months ending September 30, the average daily code contribution per engineer exceeded 45 times the annual average of 2024; the internal usage value of Claude, calculated at list price, increased more than 250-fold from January to September. None of these three sets of figures have undergone independent audits, and they respectively describe code, output volume, and internal usage metrics rather than RSI thresholds, safety levels, or cash revenue.
Policy Judgment Based on Expectations Rather Than Current Outcomes
In dealing with uncertainty, Anthropic stated that it has previously taken cautious measures such as delaying model releases and limiting initial availability, using the preliminary release of Mythos through Project Glasswing as an example. The company also indicated that it is considering updating its Responsible Scaling Policy to maintain human control in frontier model development. These arrangements constitute existing practices disclosed by the company and policy directions under consideration; the filing does not state that any new policy updates have taken effect.
TechCrunch characterized this filing as Anthropic approaching recursive self-improvement capabilities; the SEC filing's wording is more narrow, pointing to future models potentially meeting the acceleration thresholds within the company's policies and again stating that current models have not been deemed to cross the line. TechCrunch's headline represents media interpretation and does not substitute for the distinction between current status and future models in the filing. Public materials do not disclose underlying measurement data, complete threshold methodologies, specific future models, or trigger times. The current confirmation is that risk disclosures and policy evaluations have changed, but this cannot be used to confirm that RSI has been achieved.