On September 17, 2026, Anthropic published a post proposing three public measurement areas for frontier-lab AI development: the extent to which AI performs artificial intelligence research and development, the oversight of AI agents, and the allocation of computing resources. The post also disclosed company measurements from August and a one-week July snapshot, presenting a dated account of how Anthropic said AI was being used inside its own R&D processes.
Collaboration, AI-led work and autonomy
Anthropic reported that Claude had “led” 26% of its AI R&D work in August. In the company’s definition, “led” meant that Claude completed most of a task end to end from a high-level prompt while a human supervised the work. Anthropic also reported that more than 90% of the measured work was at or above its “AI collaborates” level.
Those categories are not interchangeable. The reported 26% figure describes work in which the model handled most of a task under human supervision; the higher-level collaboration figure uses a separate category; and full autonomy is a further threshold. Anthropic said Claude was not fully autonomous for any measured subset of its AI R&D as of August. The release therefore records substantial reported AI participation in R&D without establishing that Claude independently conducted the work or that the company had reached autonomous AI research.
Anthropic separately reported that approximately 30,000 agents were doing research and engineering work at any one time on its most-used internal platform during August. That figure is limited to the identified platform and period. It is not, on the evidence reviewed, a count of fully autonomous R&D systems; the same disclosure said no measured subset of the work had reached full autonomy.
For compute allocation, Anthropic reported that about 6% of the computing resources assigned to AI R&D during a July 13–20 snapshot went to safety work. The company described the estimate as conservative. Because the figure comes from a single week, it is a snapshot of the reported allocation rather than a demonstrated long-term pattern or a sector-wide measure.
What the evidence establishes
Anthropic’s methodology used Claude-based agents, judges or classifiers for parts of the task-discovery, evaluation and compute analyses. The post also described best-effort internal labels and limited sampling. Anthropic said that common measurement methods and independent third-party verification would be needed for meaningful comparisons and outside trust.
The Associated Press separately reported on the announcement on September 18, repeating the numerical disclosures as claims made by Anthropic. That provides independent confirmation that the announcement and its attributed claims were reported publicly. It does not independently audit the percentages, agent count or compute allocation. The reviewed evidence therefore confirms the publication and subsequent coverage, but not the underlying measurements themselves.
Why this enters the Record
The significance of the release is not that it settles how quickly AI is changing research and development. It is that Anthropic publicly proposed a vocabulary and supplied numerical markers for three questions that are otherwise difficult to compare: how much R&D work AI performs, how many agents are active in a lab’s internal workflows, and how compute is divided between development and safety work.
That makes the disclosure a timestamped baseline. Future comparisons could be more meaningful if other organizations adopt common definitions and independent evaluators test the measurements. At the same time, the limitations are part of the historical object. The figures came from an interested company’s internal systems, relied partly on model-assisted analysis and best-effort labels, and were not independently validated in the reviewed record.
What remains unknown includes whether the reported percentages would hold under a common external methodology, how the safety allocation changes over longer periods, and how Anthropic’s figures compare with other frontier labs. The evidence supports a narrower conclusion: on September 17, Anthropic announced a proposed measurement framework and reported that AI had taken on a significant but, by its own account, still supervised and non-autonomous role in parts of its R&D.