Anthropic says that in August, Claude was directing around 26% of the research work that builds the next model, against less than 1% in February. More than 90% of that work is said to involve the model as a collaborator or more. Around thirty thousand agents are said to be running simultaneously across the company, and the monitoring systems are said to block roughly one action in 47,000. These figures arrive six days after the head of the same company asked the industry to slow down.
We described the four levels of self-improvement in theory. This is the first time a lab has published figures on its own.
What the figure measures, and what it does not
Precision matters here, because "26% of research" lends itself to misreadings.
The important word is directs. It does not mean the model has the ideas in place of the researchers. It means it steers the execution of part of the work: launching experiments, analysing results, writing evaluation code, iterating on configurations.
This is exactly what we described regarding the mathematical proof obtained with models: the human keeps the direction and delegates the endurance. Machine learning research is particularly suited to this, because it largely consists of trying many things and measuring what works.
So the figure does not say that models design themselves. It says that the share of laboratory work they execute has been multiplied by twenty-six in six months.
Why the trajectory matters more than the level
Going from less than 1% to 26% in six months is the kind of curve that makes projections difficult.
If the progression continues at the same pace, the figure passes 100% before next summer, which makes no sense — so it will slow. The question is where.
Two ceilings are plausible. A technical ceiling: there remain parts of the work that these systems do badly, notably the generation of new hypotheses, where scores remain low. And an economic ceiling: thirty thousand simultaneous agents are expensive, which the 517 billion in compute commitments makes concrete.
Monitoring systems would block roughly one action in 47,000. This figure is valuable because it finally provides the denominator we had been calling for for months. We wrote about the inventory requested by Congress that an incident recounted tells us nothing without knowing the volume over which it occurs. One action blocked in 47,000 can be read as very reassuring or very alarming depending on what you multiply it by: across thirty thousand agents running continuously, that amounts to a great many interventions per day.
The act of publishing
It is worth highlighting what is unusual about this publication, independently of the figures.
No obligation requires disclosing the share of one's own research driven by one's models, nor the interception rate of one's monitoring systems. This is information a company normally keeps to itself.
The timing is also telling: six days after an essay asking the industry to slow down, the same lab publishes the figures that document why it is asking that question. It is consistent, and it gives the argument a factual basis that essays of this kind generally lack.
The usual caveat applies: these figures are self-reported, unaudited, and the scope of what counts as "driving" is defined by the one doing the counting.
What this changes in the debate
This dashboard gives concrete purchase to a discussion that had until now been held in the abstract.
The case for slowing down rests on the idea that capabilities are advancing faster than the capacity to evaluate them. A figure that has multiplied by twenty-six in six months for the share of automated work makes that argument measurable rather than rhetorical.
It also makes more tangible what we described in our article on acceleration loops: when a tool takes part in building its successor, the pace ceases to be set by the number of researchers available.
What to take away
This is the most useful publication we have seen in a long time, precisely because it provides denominators rather than anecdotes.
What will need watching is not the 26% figure, but its derivative: where the curve flattens out, and why. If it flattens out at a technical limit, that will teach us something about what these systems cannot do. If it flattens out at an economic limit, that will teach us that the decisive factor was never intelligence but electricity.
And above all, other labs will need to publish the same indicators. An isolated figure does not allow comparison, and that is precisely what the debate needs.