Ing. Calogero Bono
Anthropic has published its first measurements of how much of its own AI research is already carried out by Claude. At the same time, in the revised S-1/A filing submitted to the SEC on November 2, 2026, it wrote that it may soon meet its own definition of AI-driven accelerated progress. It matters because a major company is declaring numbers about its internal process and admitting that, in the past, it failed to notice significant accelerations while they were happening.
A few terms before we get into the facts. RSI (recursive self-improvement) is recursive improvement, a model that helps create its own successor. AI R&D is research and development on AI. AL is Epoch AI's automation scale, from AL0 (no AI involvement) to AL5 (full autonomy). S-1 is the document filed with the SEC before a stock market listing. Compute is the processing power used to train and run models.
What did Anthropic measure about AI that develops AI?
In the official post the company proposes three measurement tools. The first concerns how much AI builds its own next version compared with humans. The second concerns the ability to supervise and stop the actions of AI agents on internal systems. The third concerns the resources that fuel the development of more capable models.
Anthropic built a prototype called the Anthropic R&D Automation Index. It catalogs every type of research work done in the company, assesses how automated each one is and aggregates the scores.
How much research work does Claude already do?
According to Anthropic, as of August 2026 Claude does not operate fully autonomously on any measured subset of research work. The model "leads" 26% of that work, meaning it completes most of the task from a high-level instruction while a human supervises. The share of work at least at the "AI collaborates" level exceeds 90%.
These are figures declared by the company, not verified by third parties. Anthropic itself points to two obstacles to comparison across labs. There is no common methodology, and the evaluator is the company's own model, which could err like the one being checked.
How are AI agents being controlled?
As of August 2026 there were about 30,000 agents working simultaneously on research and engineering, on the most used internal platform. The measurements cover only that platform. Anthropic proposes three indicators.
- Coverage, the share of actions that pass through a monitor before or after execution.
- Review latency, the time between action and check, first automatic then human.
- Escalation rate, the share of activities blocked, redirected or flagged.
The company states that individual agents rarely make mistakes, but with millions of agents even rare events become frequent. For those using agents in their workflows, the issue is close to that of jailbreaks and prompt injection.
What does the S-1 say about RSI risk?
As reported by TechCrunch, and picked up by HDblog, in its risk factors Anthropic writes that in the past it did not identify a strong acceleration until after it had occurred. It adds that internal estimates could lead it to meet "imminently" the definition of dramatically accelerated AI-attributable progress in its Responsible Scaling Policy.
No current model has been identified as beyond the threshold. The company says it does not plan to use Claude in a fully autonomous way, without meaningful human oversight, and is considering policy changes to maintain human control.
Limits to know
The automation figures come from a prototype index, calculated with the company's own models and without external verification. The agent measurements cover a single platform. The compute usage analysis covers one week, from July 13 to 20, and the official source available to us cuts off before showing its results, so we do not report them. Third-party observers that Anthropic says it plans to bring in remain to be seen.
What changes for those working in digital
For a business or developer, the signal is that AI already writes much of the code and analysis of the people building it. It is worth not treating assistants as infallible tools and always planning for human review. If you integrate models via API, expect rapid updates and behavior that changes between versions. European regulation, described in the EU AI Act guide, also hinges on transparency about models.
Our take
Publishing metrics on your own process is a useful step, and a rare one. But it is self-assessment, and an S-1 is also a marketing document, as TechCrunch points out. The value will depend on independent verification, not on the percentages. The admission of having missed past accelerations stands out. If true, the problem is not only speed, but the ability to notice in time.
Frequently asked questions
What is RSI?
It is autonomous recursive self-improvement. A model contributes to building a more powerful version of itself, which speeds up the next generation. According to Anthropic's definition, the extreme case is a model that builds its successor in full autonomy.
Is Anthropic's AI already building itself?
Not entirely. Anthropic states that Claude does not operate fully autonomously on any measured subset. However, it leads 26% of that work and collaborates on over 90% of research work.
Does Anthropic say RSI is imminent?
It says it could soon meet the policy's definition of AI-attributable accelerated progress. No current model has been identified as beyond the threshold, and the company is still evaluating the data.
Who checks these numbers?
For now, no external party. Anthropic plans to bring in independent evaluators from several organizations, with access to internal processes and data. METR has previously tested its offline monitoring.
Why are some people skeptical?
Because the S-1 serves both to inform and to promote the company. Saying it is close to a self-building model can also serve an image function, as TechCrunch notes.
External verification and human control rules, the signals to follow
The coming months will show whether these numbers become a standard or remain an isolated initiative. Third-party verification and any policy changes will matter.
- Check whether the promised independent observers arrive and what data they will be able to see.
- Follow updates to the Responsible Scaling Policy on maintaining human control.
- See whether other labs publish comparable measures using the same metrics.
