Claude AI Leads 26% of Anthropic's R&D Work

Anthropic said Thursday that its Claude AI model now leads 26 percent of the company's artificial intelligence research and development work, up from 1 percent in March. The announcement, reported by the Associated Press and by Reuters, marks one of the most detailed public measurements yet of how quickly AI systems are starting to build their successors.

The figures come from a new internal tool Anthropic calls the R&D Automation Index. The company said it will publish the measurements regularly to show outsiders how quickly AI is building the next generation of the technology, and it urged other AI developers to share similar metrics so the numbers can be compared over time and potentially across labs.

What "leading" the research means

When Anthropic says its Claude AI model "leads" a quarter of its R&D, it means the model can complete most of a given task "end-to-end from a high-level prompt" while still under human supervision, according to the Associated Press. The model is not working completely autonomously in any part of the work measured.

The scale Anthropic used comes from Epoch AI, an independent nonprofit that tracks the technology. It runs from AL0, meaning no AI involvement at all, to AL5, meaning a system works fully on its own without human input. For the 26 percent of work where Claude leads, the model operates at the AL4 level: it takes on most of a task after receiving a high-level instruction, with a human staying in the loop to supervise.

The remaining picture is one of close collaboration. About 90 percent of the company's research and development is done in collaboration with Claude, which Anthropic said means the model does large chunks of work under close human direction. That covers model training, reinforcement learning, evaluation, and engineering.

The scale of Anthropic's agent operation

The numbers behind the shift are large. Anthropic said about 30,000 AI agents were doing research and engineering work on its main internal platform at any one time in August, and more than one billion agent decisions were tracked that month, according to Reuters. Every action the agents take is screened before it is carried out, and roughly one in 47,000 decisions was blocked.

Anthropic also disclosed that about 6 percent of its AI research compute went toward safety work during a week in July. For compute used specifically in AI-driven research, the figure was closer to 12 percent. The company shared the numbers as part of the same report, which frames safety research as a distinct and tracked category inside its growing automation footprint.

The contribution has risen fast. On the Epoch AI scale, Claude's share of work it leads climbed from 1 percent in March to 26 percent by August, and was below 1 percent as recently as six months earlier, in February 2026.

Why Anthropic is publishing the numbers

The announcement landed in the middle of an industry-wide argument about whether frontier AI development should slow down. Anthropic CEO Dario Amodei published an essay on September 12 titled "We Must Pace the Frontier," calling on AI companies to slow the rate at which they improve their most capable systems. According to the Associated Press, Amodei has been one of the leading voices in that push, which has drawn support from OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis, and Elon Musk.

In a blog post accompanying the announcement, Anthropic said that as leaders consider pacing AI's development, "we should do everything possible to minimize the gap between what frontier labs know and what the public knows." The company wrote that this means better measuring AI development, reporting on it publicly, and giving society a chance to decide how to use the information.

The company also said that models accelerating their own development could make it "more challenging for humans to understand or control these systems." Sharing these metrics, Anthropic argued, could lead to a better understanding of how close leading AI labs are to reaching recursive self-improvement, or a model's ability to autonomously build its successor.

The disclosure follows a parallel move by OpenAI. On Wednesday, OpenAI said it would begin regularly publishing reports on unexpected or unauthorized AI behavior, and it disclosed six reports of unexpected or concerning model behavior, reported by Reuters. The two announcements point to a new pattern: the leading labs are starting to publish regular transparency data on how their systems behave and how those systems are built. GenZ NewZ has also tracked Anthropic's 2026 trajectory, including Claude Opus 4.7 and its agent tools.

The numbers give the public a baseline: Anthropic said the measurements will be published on a regular schedule, and it asked other AI developers to adopt a shared public methodology. If other labs follow, outsiders will be able to track, month by month, how much of each new model was built by the previous one.