Claude Leads 26% of Anthropic’s AI R&D. Is Recursive AI Starting?

Anthropic has put a number on a transition that AI labs usually describe only in broad terms: Claude now leads 26% of the company's AI research and development work. More than 90% of measured work has reached at least the "AI collaborates" level, while roughly 30,000 internal agents are running research and engineering tasks at any given moment.

Those figures come from Anthropic's new R&D Automation Index, published with a detailed account of how the company measures automation, monitors agents and allocates computing power. They do not show a system autonomously designing, training and releasing its successor. Anthropic says none of the measured work has reached that level.

They do show something more consequential than a coding-assistant benchmark. Claude is completing substantial parts of the process that produces future Claude models. The feedback loop has begun under human supervision, and Anthropic is trying to measure its speed before it becomes difficult to observe from outside the lab.

The sharper question is whether oversight is scaling as quickly as the work. Anthropic reports more than one billion agent decisions in August, a monitoring system that blocked about one in every 47,000, and a human-review funnel that receives roughly 50 high-priority cases per week. Against that activity, about 6% of the compute used for AI R&D went to safety work during one sampled week.

Claude is helping build the next Claude

Anthropic borrowed a six-level automation scale proposed by Epoch AI. AL0 means no AI involvement. At AL3, AI collaborates on large portions of a task under close human direction. At AL4, AI leads: a person gives a high-level prompt, the system handles most of the task end to end, and a human supervises the result. AL5 is full autonomy, with no human required in the loop.

The 26% figure is the share of Anthropic's measured R&D work at AL4. It does not mean Claude performs 26% of every researcher's job or that Anthropic has replaced a quarter of its research staff. The index weights categories by estimated person-time and applies an automation rating to each part of a frozen task map.

That map is unusually granular. Anthropic sampled 20% of staff each week during July across departments involved in model R&D, extracted about 15,000 tasks from work records, and organized them into a tree with 542 nodes and 378 leaves. Examples include evaluation-platform defect diagnosis, reinforcement-learning sandbox policy, and serving-incident postmortems.

The result matters because it measures production work inside a frontier lab rather than performance on an external test. Benchmarks tell us what a model can do in a controlled setting. This index asks how much responsibility the model is already carrying in the workflow that builds the next generation.

Is this recursive self-improvement?

Not yet. Recursive self-improvement, in its strongest form, means an AI system can build a more capable successor without a human directing or approving the important steps. Anthropic explicitly says Claude has not reached full autonomy in any measured subset of its AI R&D.

But waiting for AL5 before paying attention would miss the economically important phase. A supervised system can still compress research cycles, run more experiments, repair infrastructure and free specialists to move to harder problems. If each model generation makes the next generation arrive faster, the loop can accelerate well before humans disappear from it.

That is why the move from assistance to leadership deserves scrutiny. The meaningful threshold is not a dramatic moment when a model suddenly "builds itself." It is the point at which human researchers stop being the rate-limiting step across enough of the development pipeline.

Anthropic's own August 2026 risk report treats automated R&D as a distinct threat model. TECHi previously examined the company's decision to raise its estimate of misalignment risk. The new automation data supplies an operational measure for a debate that had largely depended on forecasts.

Thirty thousand agents create a different oversight problem

Anthropic says approximately 30,000 agents were working concurrently on research and engineering in its most-used internal platform as of August. Every action in that platform passes through an online monitor before execution and is ingested by an offline monitor afterward.

The online system reviewed more than one billion decisions during August and blocked 0.002%, or about one in 47,000. The offline system flags roughly one or two transcripts per thousand, then passes about 100,000 transcripts per week through additional classifiers. Around 50 of the highest-priority cases reach human reviewers.

A low block rate can mean the controls are working and the agents usually behave as intended. It can also mean the monitors are searching a vast decision stream for rare failures whose cost may be high. Coverage is therefore only the first test. Detection quality, review speed and the consequences of false negatives matter more as agent authority expands.

There is a visible trade-off in Anthropic's design. Online monitors act within seconds, but people may take up to a week to review blocked actions. Offline monitoring gives broader behavioral context, yet it is retrospective. That split makes sense for reversible work; it is less comforting if an agent can make a consequential research or infrastructure decision before a pattern becomes legible.

The company has already seen the limits of automated safeguards in product settings. TECHi's report on Claude Code's planned auto mode described an 11% miss rate in an internal permission-control test. Internal R&D agents operate in a more constrained environment, but the lesson is the same: a monitor's coverage percentage does not reveal its recall.

The safety-compute number needs context

Anthropic measured compute allocation from July 13 through July 20. It classified about 6% of compute used for all AI R&D as safety work. Within AI-driven AI R&D, the safety share was about 12%.

Those numbers are conservative by the company's definition. Work that advanced safety and capabilities equally was assigned to the capabilities side, and separate safeguard classifiers were not included. Anthropic also notes that safety research can be labor-intensive without consuming as much compute as large training or evaluation runs.

Even so, the ratio is useful because it can be tracked. If AI-led research expands from 26% toward a majority of the development process while the safety share remains flat, oversight may be losing relative ground. If the safety share rises, outsiders still need evidence that the extra compute is improving monitoring, evaluations and incident response.

Anthropic's Responsible Scaling Policy sets capability thresholds and risk controls, but the new report points toward a simpler public ledger: the share of R&D led by AI, the share of agent actions monitored, the time to human review, and the share of compute devoted to safety. Reported consistently, those measures could reveal acceleration before a new model launch makes it obvious.

What the audit does not prove

The index is a self-audit. Anthropic used Claude agents to reconstruct staff work, classify tasks, organize the task tree and judge automation levels. A separate Claude judge agreed exactly with human raters 59% of the time. Human raters agreed exactly with each other only 35% of the time, and model-human ratings fell within one level 97% of the time.

That is evidence that the categories are usable. It is not independent validation. A model evaluating a workflow powered by related models may share blind spots with the system it judges. The frozen July task map may also miss new forms of research created by automation, especially work that no longer resembles a human process.

Anthropic acknowledges both problems and says it plans to give independent evaluators access comparable to its internal risk teams. That commitment should become the standard, not an optional credibility boost. The major frontier labs now make claims about safety, productivity and automation that cannot be tested from public interfaces alone.

The policy debate has begun to move in that direction. TECHi recently covered the attempt by major labs to write a shared AI code of conduct. Common definitions for AI-led research would make those commitments measurable across companies instead of leaving each lab to choose its own denominator.

The signal to watch next

The most important follow-up is not whether Anthropic publishes a higher 26% figure next month. It is whether the company publishes the same basket, weighting method and uncertainty range often enough to show a trend—and whether another evaluator can reproduce it.

Three changes would be meaningful. A rapid rise in AL4 work would show Claude taking operational control of more research. Any move to AL5 would indicate a task no longer requires human initiation or approval. A widening gap between automation and safety resources would show governance falling behind the pace of development.

For now, "recursive AI" is too strong a label. "Pre-recursive acceleration" is more accurate: Claude is already helping build Claude at a scale large enough to measure, while people still choose goals, supervise consequential work and decide what ships.

That distinction should not make the finding feel small. The era of AI contributing to its successor is no longer a scenario in a forecasting paper. Inside one frontier lab, it is a production metric.

Featured image: “Front of server racks at NERSC,” released under [CC0 via Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Front_of_server_racks_at_NERSC.jpg). TECHi crop, color grade and watermark.