On November 2, 2026, Anthropic disclosed in an amended registration statement (S-1/A) filed with the U.S. Securities and Exchange Commission (SEC) that AI-driven acceleration of research and development by future models could soon reach the threshold set out in its safety policy. The statement rests on the company's internal estimates and projections for future models. It does not mean Anthropic has determined that current models have already crossed the threshold.
The issue is not whether today's Claude can rewrite its own program. It is how much of the R&D for next-generation AI models can be handled by AI, and how much faster AI progress becomes as a result. According to Anthropic, Claude wrote more than 90% of the code added to its internal codebase in September 2026. But automating code writing is not the same as autonomously carrying out research as a whole. The disclosure highlights two questions: how far AI is accelerating AI R&D, and whether humans can keep track of and manage that progress.
The Condition Anthropic Is Watching: AI Progress Running at Twice the Pace
Anthropic's voluntary safety framework, the Responsible Scaling Policy (RSP), sets two thresholds for AI R&D capabilities. According to RSP v3.4, effective July 8, 2026 (pages 8–10), one is that AI can replace all of the company's research scientists and research engineers at a competitive cost. The other is that automation of AI R&D dramatically accelerates progress in overall AI capabilities.
Two benchmarks for the pace of capability improvement are used to assess the latter:
- The pace of progress expected if AI had not contributed significantly to R&D
- The fastest sustained pace of progress actually observed during a period when AI did not contribute significantly to R&D (evaluated across at least three model generations)
Anthropic sets the threshold at the point where the pace of progress is observed or projected to be twice as fast as both benchmarks, and where a substantial portion of that acceleration can be attributed to automation of R&D. It is also a condition that, if current technical trends continue, further acceleration is likely. Speeding up development simply by adding more researchers or computing resources is distinguished from AI-driven R&D acceleration.
The key point is that doubling researchers' work efficiency is not the same as AI model capabilities advancing twice as fast. The former concerns the time and effort required for individual tasks; the latter asks how quickly R&D improves the capabilities of new models. Even if AI writes large amounts of code, the RSP threshold cannot be judged met without verifying that this output translated into improved capabilities in the next-generation model.
AI-driven R&D acceleration must also be distinguished from fully autonomous "Recursive Self-Improvement (RSI)." In an explainer from the Anthropic Institute, RSI is described as a state in which AI autonomously designs and develops successor models and repeats that improvement. The company says it has not reached this stage and does not believe it will necessarily do so. AI accelerating R&D under human supervision does not mean AI can be entrusted with everything from setting research goals to completing a successor model.
Claude Wrote Over 90% of the Code, but That Is Not the Same as Automating Research
The figures Anthropic has published show that AI involvement in R&D is expanding rapidly. However, the share of code written, the automation rate of research tasks, and employees' perceptions each differ in what they measure and how.
In a pilot automation index covering R&D work in August 2026, the share of work led by AI was 26%. Meanwhile, the code contribution rate in September exceeded 90%. The gap between the two cannot simply be read as progress toward autonomy, or as the share of work AI cannot handle.
The main indicators Anthropic has published are summarized below.
| Indicator | Period | What was measured and how | What can be read from it |
|---|---|---|---|
| Over 90% code contribution | September 2026 | Share of contributions to the internal codebase written by Claude | Not the automation rate of research overall; the share reviewed by humans is also unknown |
| 26% AI-led | August 2026 | Share of R&D work falling under AL4 in a prototype index that classifies and weights tasks | AI handles most of the work but requires human supervision |
| 51% employee perception | August 2026 | Median of employees' answers on the share of work AI handled in developing frontier models in early 2026 | A subjective estimate, not measured work hours |
| RSP 2x pace of progress | Effective July 8, 2026 | Comparison with the pace of progress without significant AI contribution to R&D | A criterion in the safety policy, not an achieved result |
In its explanation of the R&D automation index, Anthropic classifies tasks into six levels, AL0 through AL5, according to the degree of AI involvement. AL4 is the stage where AI carries out most of a task given broad instructions, with humans supervising. AL5 refers to AI working fully autonomously without human intervention. As of August 2026, no AL5 operation was found in any of the R&D work measured.
The index was built by using Claude to classify R&D tasks and assess their automation level. The assessments were checked against judgments by staff, and human work records were used to weight the aggregation. However, because Anthropic's own AI is used in the evaluation, the model being evaluated and the model doing the judging may make the same kinds of errors. The company acknowledges that comparing results across AI developers will require common measurement methods and third-party verification.
The volume of code written also cannot capture the judgment R&D requires. Anthropic's explainer on RSI says that AI performs well at running and optimizing experiments with clear goals, but still struggles with choosing research goals themselves and deciding which direction research should take. Lines of code also measure quantity rather than quality of output, which can lead to overestimating productivity gains. The fact that AI handled implementation does not support the conclusion that it also made the decision to choose that experiment or evaluated the results.
Anthropic Also Disclosed It Recognized R&D Acceleration After the Fact
One notable point in the SEC filing is Anthropic's disclosure that it did not recognize AI-driven R&D acceleration early enough. According to the company, assessing AI's effect on development speed relies heavily on "lagging indicators" that can only be confirmed after changes have occurred. As a result, it has at times recognized that development speed had significantly accelerated only after the acceleration had taken place.
This is not an admission that AI ran out of control or that humans lost control. The explanation is that a gap can arise between actual progress in R&D and the point at which that progress is measured and recognized. Assessing contributions accurately becomes especially difficult when the latest or unreleased models are used in R&D. If AI capabilities advance faster than safety measures and oversight are put in place, development could proceed without risks being adequately understood.
To address this, Anthropic has also put monitoring in place for the AI agents used in R&D. According to its report on measuring R&D, as of August 2026 the internal platform the company uses most combines online monitoring, which checks agent actions before execution, with offline monitoring, which analyzes records after execution. The former aims to block dangerous operations in advance, and the latter is designed to detect problems such as those that appear across multiple actions.
However, being subject to monitoring is not the same as problems being reliably detected. Evaluating a monitoring system requires distinguishing three things:
- Coverage: How much of the AI agents' behavior can be monitored
- Time to review: The time from when an action occurs until automated checks or human review are complete
- Detection and response capability: How well problematic actions are detected and lead to necessary intervention
Even if a high share of actions is monitored, opportunities for intervention can be lost if finding problems or human review takes too long. Conversely, even if review is quick, other risks remain if the range of problems the system can detect is narrow. This is a general challenge in evaluating monitoring systems, and does not indicate that a specific monitoring failure has occurred at Anthropic.
Moreover, a mechanism for monitoring individual AI agents' actions and one for measuring acceleration of AI R&D as a whole serve different roles. The former checks whether work in progress has any problems; the latter evaluates how much faster model capabilities have improved across multiple generations. Monitoring agent actions appropriately does not necessarily mean the acceleration of R&D is being accurately captured.
How Can Progress Toward Recursive Self-Improvement Be Verified?
Anthropic does not believe the current Claude can fully autonomously and safely design and build next-generation frontier models without meaningful human oversight or understanding. The SEC filing also said the company is considering updating the RSP to maintain human control as AI R&D automation advances. It is a stance that distinguishes between being technically able to automate research and being able to entrust that research safely.
If the company later determines that it has reached the R&D capability threshold, verifying that judgment will require concrete evidence. Over what period was the improvement in AI capabilities measured, and how was it checked against the RSP's two benchmarks? How were the effects of increased computing resources and personnel separated from the contribution of R&D automation? Without this information being made public, it will be difficult for outsiders to assess whether AI-driven R&D acceleration actually met the threshold.
Third-party verification will also be a focus going forward. The Frontier Safety Roadmap that Anthropic updated on October 19 lays out a plan to publish, with a target date of December 15, 2026, a policy proposal for embedding independent third-party evaluators inside AI developers. It also aims to publish, on January 15, 2027, a policy framework addressing the pace of AI development itself. Both are currently targets, not systems already required or proven effective.
In evaluating such efforts, it will matter which models and internal data third parties can access, whether they can independently verify R&D acceleration, and whether the conditions under which humans can halt or modify development when necessary are clearly defined. Anthropic's approach to the threshold it set suggests that AI R&D automation may be entering a new stage, but it does not mean fully autonomous recursive self-improvement has been achieved. What will be asked from here is not only how much AI can speed up R&D, but whether humans can recognize, verify, and control that acceleration at the right time.