Anthropic reports Claude leading 26% of its model R&D on a definition it wrote itself

A researcher writes a paragraph of instruction, hits return, and reads what comes back. Under Anthropic's scale, that exchange is what counts as Claude leading: the model completes most of a task end-to-end from a high-level prompt while a human supervises ^[3]. The score comes off Epoch AI's automation rating scale, which grades how much of a task the model performed ^[9]. The disclosure does not say what the 26 percent is a share of. A reader cannot tell whether a quarter means a quarter of the hours, a quarter of the tickets, or a quarter of a sampled task list.

The 26 sits inside the 90. Anthropic counts the led work as part of the wider collaborative figure ^[6], so subtracting one from the other leaves about 64 points of research and development where Claude does large chunks and a human directs each one ^[19].

The Associated Press reported that the led share was none in February and reached a quarter in August, six months later ^[7]. That averages roughly 4.3 points a month if the climb were even ^[20]. Engadget describes the same chart as plotting Claude's automation level since August 2025 ^[14], so the two accounts put different windows on one series.

The disclosure landed while leading figures in AI, Anthropic chief executive Dario Amodei prominent among them, were calling for a slowdown over safety concerns ^[22]. Anthropic asked other developers to publish similar metrics regularly under a public methodology, so numbers could be compared over time and potentially across labs ^[12], and said any frontier model maker could reproduce the index from its own data with third-party validation ^[13]. "We should do everything possible to minimize the gap between what frontier labs know and what the public knows," the company said in a blog post ^[15]. It also said models accelerating their own development could make it "more challenging for humans to understand or control these systems" ^[16]. According to the AP, it was unclear from the disclosure how close Anthropic believes it is to recursive self-improvement, which the company describes as a model's ability to autonomously build its successor ^[17][21].

For anyone who has to report their own AI adoption to a leadership team, the definition is the reusable part. Two questions sort the work: whether the model started from a high-level prompt or from a specified subtask, and whether a human read the output before it shipped. High-level prompt with a human reading is Anthropic's "leads". Specified subtask with a human reading is the rest of its 90 percent. High-level prompt with nobody reading is the cell Anthropic says is empty, because Claude is "not operating fully autonomously for any measured subset of AI R&D work" ^[4].