Anthropic Says Claude Leads 26% of R&D for Its Next Models—but Not Autonomously

Anthropic disclosed Thursday, September 17, that Claude now “leads” 26% of the company’s measured artificial intelligence research and development work. That means the model can handle most of a qualifying task from a high-level prompt while a person supervises—not that Claude is independently designing, training and releasing its own successor.

The figures, based on Anthropic’s internal work as of August 2026, also show Claude working at the company’s “collaborates” level or higher across more than 90% of AI R&D. At that level, the model performs substantial portions of a task under close human direction.

Just as important, Anthropic reported no measured category of AI R&D in which Claude operates fully autonomously, without a person in the loop.

The disclosure is significant because it gives the public an unusually detailed view of how a frontier AI laboratory uses its current models to develop future ones. It also arrives during an intensifying debate over whether AI-assisted research could accelerate into recursive self-improvement faster than companies, monitors and regulators can respond.

What Anthropic’s 26% figure actually means

Anthropic applied a six-level automation scale, adapted from work by the independent research organization Epoch AI, to different categories of model development. The scale runs from no AI involvement to full autonomy.

The 26% number represents the share of Anthropic’s weighted R&D work that its index placed at AL4, or “leads.” It is not the percentage of employees replaced by Claude, the percentage of all code written by Claude or a direct measure of how much intellectual credit the model deserves.

Similarly, the figure showing collaboration on more than 90% of work should not be read as Claude independently doing 90% of Anthropic’s research. It means more than 90% of the measured work reached at least AL3, a broad category that still includes close human direction.

The change over time is nevertheless striking. Anthropic’s chart puts the share of AI R&D led by Claude at under 1% in February 2026, rising to 26% by August. The new development is the company’s public release of that measurement, not a claim that Claude suddenly crossed the threshold this week.

How Anthropic calculated the number

Anthropic built what it calls the R&D Automation Index by first mapping the kinds of work involved in its model-development process. For each week in July, it randomly sampled 20% of staff in the departments participating in that process.

A Claude research agent examined internal records, including Slack activity and other documentation, to identify roughly 15,000 granular tasks. Claude then organized those tasks into a hierarchy containing 542 categories, including 378 detailed end categories covering work such as evaluation-platform fixes, reinforcement-learning infrastructure and incident reviews.

Anthropic used a separate Claude judge to study the available evidence and assign an automation level to each category. The company weighted the categories using estimated staff time, giving more influence to work that occupied a larger share of employees’ schedules.

Human specialists also rated the automation of the areas they worked on without first seeing the model’s conclusions. Anthropic said the AI judge’s ratings were within one automation level of the human ratings 97% of the time, though exact agreement was lower and the company acknowledged genuine ambiguity at the boundary between collaboration and leadership.

This methodology provides more substance than an executive estimate or a count of AI-generated code. It also introduces several limitations:

• Claude helped build and score the index measuring Claude. Anthropic says humans checked the results, but it acknowledges that a model used as a judge could reproduce some of the same errors as the system being assessed.
• The classifications remain subjective. Reasonable reviewers can disagree about when substantial assistance becomes AI-led work.
• The task basket is frozen to a baseline. Rising automation shows Claude taking on more of the work represented in that basket, but may not capture entirely new responsibilities that employees begin performing.
• There is no comparable industry benchmark yet. Another laboratory could define tasks, human oversight and task importance differently, making direct comparisons unreliable.

Anthropic said outside verification or cross-checking by other developers’ models could make future measurements more credible. It has called on competing frontier laboratories to publish recurring figures under a shared methodology.

Why this is not recursive self-improvement

Recursive self-improvement generally describes a system autonomously contributing to a more capable successor, which then accelerates the development of another generation with progressively less human involvement.

Anthropic’s data indicates movement toward greater automation inside that loop, but it does not show that the loop has become autonomous. Based on the company’s definitions, people still provide goals, maintain supervision, review work and control whether changes are deployed.

The disclosure also does not establish that Claude can independently perform the complete model-development cycle. Anthropic did not say the model can, without human intervention:

• Choose the company’s overall research strategy or decide which successor model should be built;
• Authorize major training runs and allocate the required computing resources;
• Determine whether a model has passed safety and capability evaluations;
• Approve changes to production systems or release a model to customers;
• Detect every important problem in its own work; or
• Build and deploy a more capable successor from beginning to end.

The company’s own index places fully autonomous AI R&D at zero. The disclosure therefore supports a narrower conclusion: Claude has become a major supervised participant in the work used to build future Claude models, and that role has expanded quickly.

Anthropic is running about 30,000 research agents at a time

Alongside the automation index, Anthropic released figures describing how it monitors the agents operating on its most-used internal research and engineering platform.

As of August, approximately 30,000 agents were working on that platform at any one time. In this context, agents are parallel software instances and workflows powered by AI—not 30,000 distinct models or human-equivalent employees.

Anthropic said every action from those agents passed through an online monitor before execution and was also ingested by an offline monitoring system afterward. Across more than one billion agent decisions analyzed during August, the online system blocked 0.002%, or about one in every 47,000.

That rate should not be interpreted as proof that one in 47,000 decisions represented confirmed malicious behavior. A monitor can block uncertain or suspicious activity as a precaution, and Anthropic did not provide a comprehensive public breakdown of the blocked actions or their false-positive rate.

The offline system flags approximately 100,000 transcripts each week for additional automated classification. Anthropic said most are resolved before human review, with roughly 50 of the highest-priority cases per week escalated to people.

Those numbers reveal the operational scale behind AI-assisted research. They also show why even a low error or escalation rate matters: rare events become routine monitoring work when agents make billions of decisions.

Anthropic also disclosed how much R&D compute went to safety

For a one-week snapshot from July 13 through July 20, Anthropic classified approximately 6% of the computing power used for AI R&D as safety work. Within R&D conducted by AI agents, the safety share was about 12%.

The company described those as conservative estimates because work that advanced safety and model capabilities equally was classified as capabilities work. It also cautioned that compute is an imperfect proxy: some safety research requires extensive human analysis but relatively little computing power.

The figures are therefore best understood as a proposed transparency metric, not a complete measure of Anthropic’s safety investment. They do not represent the percentage of company spending, staff or total infrastructure devoted to safety.

Why Anthropic is releasing these numbers now

The disclosure lands after a week of unusually public disagreement over how quickly frontier AI development should proceed. Anthropic CEO Dario Amodei recently called for a coordinated slowdown and greater access for outside evaluators.

That debate gained urgency after Anthropic researcher Jacob Coxon resigned with a warning that safeguards were falling behind the AI race. Rival OpenAI has separately disclosed six incidents involving unexpected or concerning model behavior under a new reporting framework.

Anthropic argues that policymakers and the public need visibility into the production process behind increasingly capable systems—not only benchmark results released after a model has been trained.

Measures such as the share of AI-led R&D, the percentage of agent actions monitored and the allocation of research compute could eventually become triggers for additional safeguards. One possibility raised by Anthropic is a fixed testing period before a newly developed model can be used to accelerate further AI research.

What the disclosure does not tell us

Despite the level of methodological detail, several questions remain unanswered.

• Which activities account for most of the 26%? Anthropic did not provide a public task-by-task breakdown showing whether AI leadership is concentrated in software engineering, evaluations, data work, training infrastructure or research analysis.
• Which Claude versions performed the work? The published index does not provide a complete model-level breakdown of the agents assigned to each category.
• How much faster or better is the work? An automation rating does not directly measure productivity gains, quality, research novelty or the rate at which humans reject Claude’s proposed solutions.
• What successor is being developed? Anthropic did not announce a new Claude model, release date, capability target or product launch as part of the disclosure.
• Can outsiders reproduce the result? The analysis depends on confidential internal communications, documentation and workload data that independent researchers currently cannot inspect in full.

Anthropic has said it plans to embed independent evaluators from multiple organizations with access comparable to its internal risk teams. Whether those evaluators can verify the automation index—and how much they will be allowed to publish—will be an important test of the company’s transparency commitment.

What to watch next

1. Recurring updates. A single August snapshot is informative, but a consistent series will show whether the jump from under 1% to 26% continues, stalls or reverses.
2. Movement toward AL5. The most consequential signal would be any measured R&D category moving from supervised leadership to full autonomy.
3. Independent verification. Third-party access could identify whether Claude’s self-evaluation, Anthropic’s task weighting or its definitions materially inflate or understate automation.
4. Oversight capacity. Agent counts and decision volumes may grow faster than the number of cases humans can investigate, making review latency and monitor accuracy as important as coverage.
5. Comparable disclosures from rivals. A shared reporting standard would make it possible to compare the pace of AI-assisted research across Anthropic, OpenAI, Google DeepMind and other frontier laboratories.

For now, the clearest signal is speed rather than independence. Claude is doing far more of Anthropic’s model-development work than it was six months ago, including leading a meaningful share of supervised tasks. But Anthropic’s own measurements still show humans deciding what to build, overseeing the work and controlling deployment.