Anthropic AI R&D Automation Index Explained: Is Claude Building the Next Claude?

Anthropic created a catalogue of AI research and development work, including areas such as training, evaluation, research engineering and infrastructure. The lab then rated tasks using an automation scale developed by Epoch AI. The scale runs from AL0, where AI is not involved, through levels of assistance and collaboration to AL4, where AI can complete most of a task end-to-end from a high-level prompt while a human supervises. AL5 would represent full autonomy with no human in the loop.

As of August 2026, Anthropic says Claude led 26% of its AI R&D work at the AL4 level. More than 90% of the work was at or above the collaboration level. The distinction matters: leading a task under supervision is not the same as deciding the research agenda, approving deployment or operating without an accountable human.

The number is best understood as a work-distribution metric. It tells us how much of a defined basket of tasks can be delegated to Claude at a particular level of supervision. It does not tell us that Claude can build an entire frontier model alone, that every R&D task is equally automated or that the same percentage would apply at another lab.

Recursive self-improvement usually means a system can autonomously build a more capable successor, evaluate the result, and repeat the cycle with little or no human direction. Anthropic's measurement shows progress toward automation of parts of the development process, not that this complete loop has arrived.

A frontier model-development pipeline contains many coupled activities: choosing objectives, designing data, writing and reviewing code, running experiments, interpreting results, allocating compute, testing safety and making release decisions. Claude may lead some categories while humans still define goals, control access, inspect outputs, resolve failures and decide which experiments are valid. That distinction is also central to our AI agents versus workflow automation guide: flexible agents need deterministic permissions and review around them.

The trend is still strategically important because automation compounds. If AI completes more experiments per researcher, teams can test more hypotheses, fix infrastructure faster and spend more time on safety evaluations. The same acceleration can also make it harder for humans to understand why a system changed, which assumptions were embedded in the process and whether monitoring is keeping pace.

Anthropic reports three oversight concepts for its internal research agents: coverage, review latency and escalation rate. Coverage asks how much agent activity passes through a monitor. Review latency asks how quickly activity is checked. Escalation rate tracks what is blocked or sent for deeper review.

The report says approximately 30,000 research and engineering agents were active at a time on its most-used internal platform in August. It describes online monitors that act before or during execution and offline monitors that review activity after the fact. Online monitoring is suited to actions that could cause immediate, irreversible harm. Offline review can detect slower patterns, such as suspicious coordination or signs of misalignment, but it cannot undo every action.

This is a useful model for any company deploying coding or research agents. A dashboard showing 100% ingestion does not automatically mean 100% safety: the monitor may miss a behavior, flag too much noise, or route alerts to a team that cannot respond quickly. Effective oversight needs coverage, meaningful detection, a clear owner, response time targets and a tested shutdown path.

Anthropic also published a snapshot of how its compute was allocated during a week in July 2026. It estimated that about 6% of compute used for AI R&D went to safety work, and about 12% of compute used for AI-driven AI R&D went to safety. These are deliberately conservative estimates based on workload classification.

Compute allocation is attractive as a transparency metric because accelerator usage can be measured and compared over time. It can help observers ask whether safety resources grow as capability work grows. However, compute is an imperfect proxy: a safety researcher may spend days designing a test that uses little accelerator time, while a large evaluation run may consume substantial compute without covering the most important risk.

A credible report should publish category definitions, classification confidence, sampled workloads and whether shared infrastructure is counted. Independent review is especially valuable because the organization choosing the labels also benefits from presenting its safety work in the strongest light.

For an engineering team, the first lesson is to define work boundaries before calculating productivity. ‘The agent wrote 60% of the code’ is ambiguous unless the team records whether the number includes generated drafts, accepted changes, tests, debugging, review and deployment. AI R&D measurement has the same problem at a larger scale.

The second lesson is to pair output metrics with control metrics. A team that measures completed tasks but not blocked actions, review time, permission failures and rollback success is measuring speed without measuring operational risk. This is particularly important for agents that can run shell commands, access private repositories or launch expensive experiments. Our reliable LLM evaluation pipeline checklist offers the same evidence-first mindset for measuring AI systems.

The third lesson is reproducibility. Publish the task taxonomy, scoring rubric, sampling window and limitations. If the number cannot be recreated or challenged, it is closer to a marketing claim than a durable safety metric. External evaluators and periodic remeasurement can make the trend more credible.

Anthropic's disclosure is valuable because it replaces vague predictions with a starting measurement framework. Claude leading 26% of AI R&D means that supervised AI-led work is already a meaningful part of a frontier lab's process. More than 90% of work at or above collaboration shows how quickly the role of agents can spread even before full autonomy.

The number should not be inflated into a claim that Claude is independently designing, training and releasing its successor. The stronger conclusion is more practical: the development process is becoming increasingly agentic, so transparency and oversight must scale with it. Task definitions, monitoring coverage, response latency, escalation rates and compute categories give the public and developers better questions to ask.

The future of AI development may depend less on whether one model can perform one impressive task and more on whether organizations can keep the whole automated pipeline legible, reversible and accountable. Anthropic's index is an early attempt to measure that transition. Its real test will be whether the methodology remains consistent, improves through independent review and is adopted by other frontier labs.