Anthropic scores Claude as leading 26% of measured R&D with a prototype that uses Claude to judge

According to The Neuron's reading of the disclosure, Anthropic's scheme grades categories of work by how automated they have become ^[14]. Nothing sits in the highest grade. Anthropic says Claude is not fully autonomous on any measured portion of the work ^[4]. Going by the definitions, "leads" sits below full autonomy and above a looser grade where Claude completes large chunks under human direction ^[2][3]. Suppose the Claude-led work is counted inside the more-than-90% that has reached at least the large-chunks level. Then more than 64 percentage points of measured R&D sits in that middle grade ^[1]. The most common state of Anthropic's measured research is Claude doing large pieces while a person directs it ^[1].

Every one of those shares is a share of measured work ^[1]. For the 26% to describe Anthropic's research as a whole, the work that gets measured would have to look like the work that does not ^[1].

The classifier is the part I would push on at review. The prototype partly uses Claude to classify and judge the Claude-assisted work it counts ^[5]. For the 26% to carry over into anyone else's planning, Claude's labels would need to match human reviewers' labels on the same tasks, at an agreement rate someone had actually measured. I would not accept the arrangement in a test suite, where code under test writing its own pass criteria is the first thing a reviewer flags. Anthropic does deserve credit for publishing a precise definition of "leads" and stating the prototype status alongside the headline number ^[2][5].

OpenAI's figure measures something else. It counts effort: OpenAI tracks agent use, experiment volume, code production and task duration ^[13]. Taken as a share of volume, 3.1 agent workdays per human workday puts agents at about 76% of research-organization workdays ^[3]. For that ratio to say anything about speed, an agent workday would have to produce accepted research at something near the rate of a human one. OpenAI's own caveat on that point is the most useful sentence in either disclosure. The company says more code and more experiments do not necessarily mean AI itself is advancing faster, and that human researchers still set priorities, interpret results and decide what gets scaled or deployed ^[8]. OpenAI calls the measurements preliminary ^[9].

The 22-author paper is just as careful ^[10]. Its authors say the intelligence explosion they describe has not happened ^[16]. They also list what could break the loop: compute constraints, slow training runs, diminishing returns, difficult research problems, and tasks where humans remain the bottleneck ^[11]. The Neuron's assessment is that no independent evidence yet shows the labs' gains have created that feedback loop ^[15].