The Anthropic R&D Automation Index is a prototype measure of how much of a frontier AI developer's research and development is performed by AI rather than by people. Anthropic introduced it on September 17, 2026 in a post by the Anthropic Institute, which describes it as built "by cataloguing every kind of AI R&D work done at the company, rating how automated each task currently is, and aggregating those ratings" (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)). Anthropic presents it as a way to gauge how close the world is to recursive self-improvement, which it defines as "a model fully autonomously building its successor," and proposes that any frontier developer could publish comparable figures regularly.
Automation scale
The index rates work on an automation scale developed by Epoch AI, which measures "Automation Level" (AL) in six steps: AL0, no AI involvement; AL1, minimal AI involvement; AL2, AI assists; AL3, AI collaborates, doing large chunks of work under close human direction; AL4, AI leads, completing most of a task end-to-end from a high-level prompt while a human supervises; and AL5, AI operates fully autonomously with no human in the loop (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)). The scale derives from a June 2026 Epoch AI Gradient Updates post, "Toward an O\*NET for AI R&D," which proposed rating AI R&D tasks from 0 to 5 on how much current AI automates them (Source: epochai.substack.com). Anthropic's post uses the level names without numbering AL1 and AL2 individually; the numbering above follows the ordering it gives.
Method
Anthropic's construction has three parts (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)):
- Task map. For each week of July 2026, 20% of staff in each department of the model R&D loop were randomly sampled, and a Claude research agent reconstructed each sampled person's week from Slack and internal documentation. The resulting list of about 15,000 tasks was organized by Claude into a tree of 542 nodes, 378 of them leaves, which is then frozen so that every later measurement runs against the same basket of work.
- Ratings. For each node a Claude agent researches how the work is done across the company, and an independent Claude judge assigns an automation level. When a given month is rated, the agents see only evidence from that month or earlier.
- Weights. Nodes are weighted by person-time. Each sampled person contributes one unit of weight per week, divided evenly among the tasks they worked on, so that categories many people work on count for more.
Anthropic checked the judge against blind ratings from staff who own the relevant work areas: exact agreement was 59% between model and humans against 35% between humans, and ratings fell within one level of each other 97% of the time. To test whether a frozen basket misses newly created work, it compared tasks arising each month from February to July 2026 against a basket built from January 2026 data and found no rise in the number of novel tasks. It plans to rebuild the basket periodically and re-version published figures.
First reading
As of August 2026, Anthropic reported that Claude "leads" (AL4) 26% of its AI R&D work, that the share at or above "collaborates" (AL3) was above 90%, and that Claude was not operating fully autonomously (AL5) for any measured subset. A chart caption in the post puts the AL4 share at under 1% in February 2026 (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)). Implicator.ai reported the post as the "Anthropic R&D Automation Index" on September 18, 2026 (Source: implicator.ai).
Stated limitations and proposed uses
Anthropic names two obstacles to comparing the index across developers: the lack of a common methodology, and the use of a developer's own models as the judge, which "could make the same kinds of errors as the model it is checking." It proposes verification by a third party or by other developers' models, with guardrails on competitively sensitive data. It also notes room for disagreement on borderline cases between "collaborates" and "leads," and that a rising index on a frozen July 2026 basket shows the automation of the work humans did then rather than whether humans have moved onto new work. The post suggests such measures "could also become the trigger for stronger requirements, like a fixed testing window before a new model is used for further AI R&D" (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)).
Relation to other measures of AI R&D automation
The index is one of several approaches to measuring AI R&D automation proposed or published in 2026:
- Code and productivity measures. Anthropic's June 2026 essay reported the share of merged code authored by Claude and code merged per engineer (When AI Builds Itself (The Anthropic Institute, June 4, 2026)).
- Activity measures. OpenAI's September 6, 2026 disclosure measured researcher inference spend, agent-to-human workday ratios and experiment velocity, and classified agent-token growth against Epoch AI's task taxonomy (Research acceleration: The view inside OpenAI (OpenAI, September 2026)).
- Capability evaluations. Shadow evaluations give an agent the research question of an unpublished paper and have its authors grade the result (Can AI agents conduct open-ended AI research? Early evidence from two case studies (Kirgis et al., July 2026)).
- Mandated disclosure. The Self-Improving AI Monitoring Act discussion draft would direct NIST to assess trends in autonomous AI research and development; the Institute for Progress recommends that frontier companies publicly share information on AI R&D automation trends (How Should the US Prepare for Increasingly Automated AI R&D? (IFP, August 2026)); and Americans for Responsible Innovation's blueprint proposes quarterly mandatory disclosure on the automation of AI R&D (Responsible Innovation at the Frontier (ARI, August 2026)).
The Anthropic post pairs the index with two further measurements, oversight metrics for internal agents (Post-Deployment AI System Monitoring) and the safety share of AI R&D compute (Compute Governance).
Relationships
- depends-on: Epoch AI — source of the AL0–AL5 automation scale
- instance-of: AI Progress Measurement — a measure of AI R&D automation rather than of model capability
- related: Recursive Self-Improvement (RSI) — the endpoint (AL5 across AI R&D) the index is presented as tracking
- supports: Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026) — the post that introduces and first reports the index
- related: Research acceleration: The view inside OpenAI (OpenAI, September 2026), Shadow Evaluations, Self-Improving AI Monitoring Act, How Should the US Prepare for Increasingly Automated AI R&D? (IFP, August 2026), Responsible Innovation at the Frontier (ARI, August 2026)
- related: Anthropic, The Anthropic Institute, AI Transparency