Anthropic: Claude now leads 26 per cent of the work on new AI models, up from under 1 per cent in February

AI that builds better AI: that is the idea behind the singularity. Anthropic, the company behind the AI assistant Claude, is trying to measure how far this has already gone. In September the company published a first measurement, the R&D Automation Index (R&D stands for research and development). It shows that in August 2026 Claude led about 26 per cent of the work on new AI models at Anthropic, measured in working time. In February it was less than 1 per cent.

Leading means this: Claude gets an assignment in broad terms and does most of the work from start to finish, while a human supervises. Anthropic gives an example. If a program that processes data gets stuck, Claude looks for the cause, writes and tests a fix, but does not put it into use itself. An employee reads the report and decides.

On more than 90 per cent of the work, Claude helps at least substantially, under close direction from humans. According to the company, Claude does not work fully independently, with no human involved, on any part of the work that was measured.

The figures drew a lot of attention at the end of October because of a column in the American newspaper The Wall Street Journal. On 2 November there was more news. Anthropic is preparing a stock market listing. In its filing with the American financial regulator, it writes that it may soon reach a limit set by its own safety policy: the point at which AI greatly speeds up progress. After that, stricter rules for safety and testing apply. According to the company, no current model goes beyond that limit. Anthropic also writes that it only noticed earlier speed-ups in its work afterwards, and it is considering a rule that a human must approve certain decisions.

How Anthropic measures this

Every week in July, Anthropic randomly chose 20 per cent of the employees who work on new models. A Claude agent, a version of Claude that carries out tasks on its own, read their internal chats and documents and made a list of about 15,000 tasks from them. Then another Claude model gave each type of work a monthly score on a scale with six levels, from no AI to fully independent. Leading is the second-highest level.

Employees also rated their own area of work, without seeing the scores. Human and model gave exactly the same score in 59 per cent of cases, and in 97 per cent of cases they differed by at most one level. Two humans agreed exactly with each other in 35 per cent of cases, so the model's scores did not differ more than those of colleagues.

The company calls the index a first trial version. A Claude model is rating work done by Claude here, so it can make the same mistakes. The line between collaborating and leading is also not always sharp. And the list comes from July, so types of work that appeared later are not included. Outsiders have not yet checked the figures. Anthropic wants to bring in independent reviewers from several organisations for this and is asking other AI companies to publish comparable figures.

Who checks the Claude agents?

On Anthropic's most widely used internal platform, about 30,000 agents were working at the same time in August. Every action first passes an automatic check, which is mainly meant to prevent fast, irreversible damage, such as copying an AI model away. That check blocked 0.002 per cent of the more than one billion actions, about 1 in 47,000, and humans review those cases within a week. Anthropic warns that it has only been doing this for a few months and that it is not sure it sees everything.

The company also looked at how computing power is divided. In one week in July, about 6 per cent of the computing power for AI research went to safety work. For research that AI carried out itself, it was about 12 per cent. Anthropic deliberately calls both figures cautious estimates.

Where humans are still needed

In an earlier report from June, When AI builds itself, Anthropic showed where humans make the difference for now: in choosing which problem is worth the effort and which result can be trusted. The same report says that in May Claude wrote more than 80 per cent of the code that Anthropic added to its programs. In the second quarter, developers delivered eight times as much code per person per day as in 2024. Anthropic itself calls that eightfold figure almost certainly too rosy, because lines of code say something about quantity and not about quality. According to the authors, full self-improvement, where AI makes its own successor without humans, has not arrived yet and is not inevitable either.

Jack Clark, one of the founders of Anthropic and the author of that report, did bring forward his personal estimate for such full self-improvement from late 2028 to 2027. That is his own expectation, not a measurement.

This news connects to the report by the independent research organisation METR, which spent eleven weeks looking over Anthropic's shoulder. METR answered a different question: not how much work AI does, but how fast the models are improving. The estimate was about twice as fast as between March 2024 and October 2025, with a large margin of error, and AI mostly worked as a tool. Both point in the same direction, although Anthropic keeps much of the evidence to itself.

What does this mean for the road to the singularity?

Going from under 1 per cent to 26 per cent in six months is a big jump, measured by one company with a new method. It shows that at this company AI already carries a large part of the daily work on new models, while humans still choose what happens and what goes into use. That is the phase before self-improvement, not yet self-improvement itself. The next step is for outsiders to check this and for more companies to measure in the same way. Then everyone can follow how fast it really goes, and there is time for good agreements about who makes the decisions.

What is the technological singularity? Read the explanation →