- Published on
How Close Are We to AI Creating AI? Anthropic's Claude Already Handles 26% of Internal AI R&D
- Authors
- Name
- DeepHorus
On September 18, Anthropic released a methodology for measuring the R&D progress of frontier AI labs and published internal interim data. The data shows that as of August 2026, its AI model Claude has led about 26% of AI R&D work, and more than 90% of R&D tasks are at the level of AI collaboration or above. This data has also drawn outside attention to how far away 'AI autonomously developing the next generation of AI' still is.
This indicator system mainly focuses on three metrics: the extent to which AI participates in AI R&D, the supervision of AI agents, and the allocation of computing power between AI R&D and safety research. Anthropic stated that as AI capabilities improve, AI is increasingly participating in the development of next-generation systems. The public metrics are intended to give the public, third-party institutions, and governments a more intuitive understanding of the pace of frontier AI R&D. The company's CEO, Dario Amodei, has previously called for coordinated slowdown of frontier AI development; if such coordination occurs, the relevant figures are expected to change.
In terms of AI participation in R&D, Anthropic proposed a 'R&D Automation Index,' using Epoch AI's automation level system, from AL0 to AL5 corresponding to no AI involvement, AI assistance, AI collaboration, AI-led, and fully autonomous operation. As of August 2026, Claude had not achieved fully autonomous operation in any measured R&D work, but it had led about 26% of the work, and the proportion of work at AI collaboration level or higher exceeded 90%. This proportion was less than 1% in February 2026.
To establish the metrics, Anthropic sampled and organized about 15,000 model R&D tasks from internal work records, Slack, and other documents in July 2026, summarizing them into 542 task nodes, of which 378 belonged to the most granular task categories. Claude then studied how tasks were executed and the degree of AI involvement, and an independent Claude evaluation model made automation level judgments.
In terms of agent supervision, the real-time monitoring system checks behavior before agents execute actions, intercepting or adjusting dangerous operations. Analysis of more than 1 billion agent decisions in August 2026 showed that approximately 0.002% of operations were intercepted, equivalent to one interception per 47,000 operations. The post-hoc monitoring system flags about 100,000 conversation records each week for further inspection, of which about 50 with the highest priority are escalated to human review, and approximately 0.1% to 0.2% of agent activities are flagged for further review.
In terms of computing power allocation, Anthropic measured computing usage from July 13 to 20, 2026. The data shows that about 6% of computing power used for AI R&D was invested in safety research; if only AI-based AI R&D work is counted, safety research computing power accounts for about 12%. Anthropic emphasized that this is a conservative estimate, because work that simultaneously promotes both capabilities and safety was not counted, nor was the computing power for safety classifiers. The company also pointed out that safety research typically does not require as much computing power as frontier model training, so the computing power share cannot directly represent the overall level of investment in safety work.
Anthropic said that these metrics are still a prototype measurement system, and cross-laboratory comparisons face issues such as unified methodology and independent verification. The company plans to bring in independent third-party evaluators from multiple organizations, giving them access to processes, systems, and data similar to those of the internal risk assessment team, to verify safety practices and continuously monitor the metrics. Anthropic stated that publishing these measurements is intended to give the public, third parties, and governments a better understanding of the pace of AI development inside frontier labs, and it hopes to continue publishing relevant data.
