Anthropic says Claude now leads 26% of its model R&D, and asks rivals to publish the same numbers

Anthropic splits the 26% figure into two carefully separated buckets. Lead means the model can take a high-level prompt and complete most of a given task end-to-end, but still under human supervision; the model cannot yet work completely autonomously. The remaining roughly 90% is collaboration, which Anthropic describes as the model doing large chunks of work under close human direction. The difference between the two buckets is not the model's ceiling but whether a human is still inside the loop. The fact that lead went from zero to 26% in six months, stated in the same breath as an admission that the model is not yet autonomous, is not comfortable reading: a lab has spent six months handing over a quarter of what used to be human work.

The timing of the announcement matters more than the numbers. Just over a week earlier, an Anthropic researcher resigned with a dire warning about the threats the technology poses to humanity, touching off a fresh public debate about whether frontier model development should be slowed down. Anthropic chief executive Dario Amodei has been one of the loudest voices calling for a slowdown. That gives the disclosure a subtle tension: the company argues for restraint at the industry level while reporting, honestly, how far its own internal automation has gone. Anthropic's proposed answer, in the accompanying blog post, is transparency. Its wording is that we should do everything possible to minimise the gap between what frontier labs know and what the public knows, by better measuring the development of AI, reporting on it publicly, and giving society an opportunity to decide how to use this information. In other words, if the slowdown will not be agreed, at least let outsiders see.

Two numbers in the announcement are easier to skim past. The first is scale: about 30,000 agents were doing research and engineering work inside the company as of August. That order of magnitude explains why Anthropic devotes space to oversight. Only when there are tens of thousands of things to oversee does the rate at which monitoring catches misbehaviour become statistically meaningful, and however low the per-agent error rate is, multiplied by thirty thousand it becomes a daily event. The second is a commitment to embed external third-party evaluators inside the company to monitor safety efforts. Putting evaluators inside the building is a much heavier and harder arrangement than having them write reports from the outside, but it answers a specific question: once models start helping develop the next model, who decides that the curve is moving too fast.

What the disclosure does not answer may matter as much as what it does. It does not say how close Anthropic believes it is to recursive self-improvement, and it does not offer a single agreed standard for judging. What it does offer is a methodology others can copy: define precisely what lead and collaborate mean, publish on a fixed cadence, and make numbers comparable across labs. That is why the announcement closes by urging peers to publish the same metrics. Read it with one caveat: these are self-reported internal figures with no independent channel to verify them, and the line between lead and collaborate is drawn by the company itself, leaving room for interpretation. Treat it as a statement that the books are now being kept, more than as a report card.