Claude leads 26% of measured Anthropic R&D. The autonomous category is still zero

26% leads, more than 90% collaborates or leads, and 0% is autonomous

Anthropic has published an internal measurement framework for how much of its own AI research and development is performed with Claude. The most striking number is that Claude now leads 26% of measured work. More than 90% is classified at or above collaboration, while none of the measured subsets qualifies as autonomous.

The language matters. Anthropic defines a ladder of assistance. At the “leads” level, an AI system performs most of a task end to end from a high-level prompt, but a human still supervises the work. The autonomous level removes that human loop. Calling the 26% figure recursive self-improvement without that distinction would overstate the result. The system is helping researchers build systems, but people still define goals, review outputs and decide what enters production.

Fifteen thousand tasks turn a vague claim into a measurable one

Anthropic decomposed AI R&D into roughly 15,000 granular tasks organised under 542 nodes and 378 leaves. It then used internal evidence and model-based classification to estimate the level of AI participation. This is more useful than counting prompts because it connects tool use to work outcomes.

It is still a self-measurement system. Claude agents and judges help classify Claude’s contribution inside Anthropic’s own workflows. The company identifies that limitation. The task taxonomy, evidence quality, judge behaviour and threshold for “lead” all affect the result. Independent replication would strengthen the claim, especially across laboratories with different engineering practices.

The capability ladder in practical terms

Level	Typical relationship	Human role
Assists	AI handles a bounded part of a task	Directs most steps and integrates the result
Collaborates	AI and human share substantial work	Iterates, reviews and redirects
Leads	AI handles most of the task from a high-level prompt	Supervises and approves
Autonomous	AI performs the task without a human loop	No task-level supervision

Thirty thousand concurrent agents make one-in-47,000 events routine

Anthropic reports about 30,000 agents running at any one time on its most-used internal platform. At that scale, rare failures become routine operating events. The company says online monitors covered 100% of activity and made more than one billion decisions in August, blocking about 0.002%, roughly one in 47,000. Offline analysis ingests all transcripts, flags about one to two per thousand and sends roughly 50 per week for human review.

These controls show serious investment, but coverage is not the same as detection. A monitor can observe every run and still miss a harmful pattern. Its value depends on the rules, model, data and escalation process. The useful assurance metrics are therefore false-negative testing, time to review, remediation quality and whether blocked behaviour can reappear through another tool or workflow.

Monitoring figures reported by Anthropic

Layer	Reported scale	Question to ask
Online monitoring	100% coverage and more than one billion August decisions	What harmful behaviour was the monitor tested to detect?
Online blocking	About 0.002% of decisions	How many blocks were true positives and what happened next?
Offline review	All transcripts ingested	How quickly can high-severity patterns reach a human?
Flagging	About 100,000 transcripts per week	How are samples prioritised and clustered?
Human escalation	About 50 transcripts per week	What evidence supports that reduction ratio?

Six per cent of R&D compute went to safety in one measured week

For a one-week snapshot in July, Anthropic says safety work used 6% of AI R&D compute and 12% of AI-driven AI R&D compute. Those figures are useful because they expose an allocation decision: as models accelerate model development, some of that acceleration is being directed at safeguards.

The ratio alone does not show whether the safety work is effective or whether it scales with capability. A stronger public metric would connect compute to outcomes such as vulnerabilities found before deployment, incidents prevented, evaluation coverage and time to remediate.

A task count is not a contribution count until the tasks are weighted

Anthropic’s taxonomy has 542 nodes and 378 leaves, with about 15,000 granular tasks beneath it. The structure is valuable because it prevents “AI does research” from collapsing coding, experiment design, incident response and scientific judgement into one label.

The remaining problem is weighting. Completing ten routine code transformations is not necessarily equivalent to choosing one research direction that changes the programme. A model can lead a large share of counted tasks while humans retain control over the small number of decisions that determine scientific value. Anthropic’s public percentage measures automation across its task representation. It does not directly measure the share of useful discoveries, capability gain, revenue or risk attributable to Claude.

Epoch AI’s six-level rubric defines “leads” as completing most of a task end to end from a high-level ask while a human supervises, corrects course and approves outputs. “Autonomous” requires little or no human involvement. Anthropic adopts that distinction, which is why its 26% headline and zero-autonomy result can both be true.

What the 26% figure measures and what it does not

Claim	Supported interpretation	Unsupported interpretation
Claude leads 26%	Most steps in those classified tasks come from Claude under supervision	Claude produced 26% of Anthropic’s scientific value
More than 90% at collaboration or above	AI participates substantially in most measured work	Humans are unnecessary in most research
Zero autonomous	No measured subset met the little-or-no-human-involvement definition	No agent can ever act without a human
15,000 tasks	The taxonomy is granular	Every task has equal importance or risk

One billion monitor decisions make a tiny rate a large review system

Anthropic reports more than one billion online-monitor decisions during August and a block rate of 0.002%, about one in 47,000. Applied to one billion decisions, that order of magnitude implies tens of thousands of blocks. The company says humans review blocked actions within one week. It does not publish the exact decision count, precise block count or false-positive rate, so a more exact calculation would create false precision.

Offline monitoring ingests 100% of transcripts, flags about one to two per thousand and produces roughly 100,000 flagged transcripts a week. Multiple classifiers reduce those to about 50 high-priority cases for human review. That is an extreme filtering ratio. Its safety value depends on whether the discarded flags are genuinely harmless and whether the system detects behaviours it was not designed to recognise.

Coverage answers whether the monitor saw an action. It does not answer whether the monitor understood it. A stronger assurance report would publish challenge-set recall, false-negative testing, review latency by severity and the number of control changes caused by escalated cases.

Recursive improvement needs a closed loop, and this one still has human gates

The phrase recursive self-improvement suggests a system that improves the machinery used to improve itself, then repeats the process with decreasing dependence on people. Anthropic has published evidence for part of that loop: Claude assists with research and engineering used to build later Claude systems.

The published process still contains human choices at each consequential boundary. People define the task tree, select evidence, supervise work classified as “leads”, approve outputs, design the monitoring system and decide whether a change enters training or deployment. Claude also helps judge Claude’s contribution, creating a correlated measurement risk. Anthropic names self-measurement as a limitation and proposes third-party verification.

The practical threshold to watch is not whether AI touches AI research. That threshold has been crossed. It is whether a system can choose a consequential research objective, execute the work, evaluate the result, alter the next system and continue without task-level human approval. Anthropic’s own autonomous column says that threshold is not present in the measured work.

The question that exposes the gap

Which high-value model improvement entered production because an agent chose and completed the research path without task-level human approval?

The report does not identify one, and its autonomy measure says none of the measured subsets qualifies. That does not make the 26% result small. It makes it more precise: supervised agents are already an industrial part of frontier-model development, while the decisive scientific and deployment gates remain human.