Anthropic releases three metrics showing Claude’s role in R&D, agent supervision and safety compute

On September 17, 2026 Anthropic published three metrics designed to measure the pace of AI development in leading laboratories. The metrics focus on three areas: the proportion of research and development (R&D) work directly driven by the company’s flagship model Claude, the number of autonomous agents operating on Anthropic’s internal platform, and the fraction of compute allocated to safety‑related tasks.

Claude’s share of R&D work

According to the company’s own data, Claude currently leads 26 % of the measured R&D workload. In practice this means Claude receives a high‑level instruction, executes the bulk of a task from start to finish, and then hands the result over to a human reviewer for validation.

Anthropic also states that Claude contributes substantively to more than 90 % of the R&D that is measured, even though it never operates fully autonomously on any evaluated portion of work. The distinction is important: the model can handle the core of a task but final responsibility remains with human overseers.

Historical reference points show rapid growth. In February 2026 the metric for Claude‑directed work was recorded as zero. By March 2026 another measurement cited by Bright placed the figure at 1 %. The September 2026 metric therefore reflects a steep increase, though the two earlier figures used a different methodology and cannot be directly merged with the latest data.

Scale of agent supervision

Anthropic’s internal platform hosts roughly 30 000 agents that simultaneously perform research and engineering tasks. These agents are monitored by a version of Claude that classifies each piece of work according to the Epoch AI automation scale, after which humans double‑check the classifications.

Bright reported that in August more than one billion decisions were made by the system, and only about one decision in 47 000 was blocked by internal safety checks. This figure comes from Anthropic’s own internal controls and illustrates the low rate of manual intervention at scale.

Compute dedicated to safety

A snapshot taken between 13 and 20 July 2026 shows that 6 % of the total compute allocated to AI R&D was devoted to safety‑related activities. When the compute is limited to the portion of R&D performed by AI systems, the safety share rises to 12 %.

Anthropic emphasizes that all three metrics are generated using its proprietary data and models. They are intended as a starting point for industry discussion and are not yet validated by external auditors, nor do they claim to be universal measures of progress.

• Claude leads 26 % of measured R&D work
• Claude contributes to over 90 % of measured R&D tasks
• 30 000 agents run concurrently on the internal platform
• 6 % of total R&D compute, 12 % of AI‑driven R&D compute, allocated to safety

The release of these metrics signals Anthropic’s attempt to bring transparency to a field where internal benchmarks are rarely shared. By quantifying Claude’s role, the company provides a concrete data point that can be compared with other labs that are experimenting with large‑scale model‑assisted research.

For English‑speaking organisations, the practical implication is clear: the metrics offer a template for tracking how much of their own R&D pipeline is being automated, how many autonomous agents are in operation, and what proportion of compute is earmarked for safety. Companies can adopt similar measurement frameworks, adapt the automation‑scale classification, and set internal thresholds for human review based on the low‑frequency blocking rate reported by Anthropic.

In summary, Anthropic’s three metrics provide a measurable snapshot of AI‑driven research at a leading lab. While the figures are internal and await external verification, they give other firms a reference for building their own dashboards, calibrating human‑in‑the‑loop processes, and ensuring that safety remains a visible slice of the compute budget.