Anthropic says its Claude artificial intelligence system now “leads” 26% of the research and development work involved in building the company’s next generation of AI technology, offering a rare look at how quickly advanced AI is becoming part of its own development process.
The San Francisco-based AI company disclosed the figure on September 17 as part of a new set of measurements designed to track how much AI contributes to frontier-model research, how AI agents are monitored and how computing resources are divided between capability development and safety work.
As of August 2026, Claude was leading 26% of Anthropic’s AI R&D work, while more than 90% of the research was being conducted with AI involved at least as a collaborator, according to the company. Reuters reported that the share categorized as AI-led had risen sharply from about 1% in March.
The numbers provide a notable measure of how rapidly artificial intelligence is moving from being primarily a tool used by researchers toward becoming an active participant in the development of more advanced AI systems.
What Anthropic Means When It Says Claude ‘Leads’ Research
Anthropic stressed that Claude is not operating fully autonomously in any measured part of the company’s AI research and development.
The company based its assessment on an automation scale developed by independent research organization Epoch AI.
Under that scale, an AI system at the “collaborates” level can complete substantial parts of a task while working under close human direction. At the higher “leads” level, the AI can handle most of a task from beginning to end after receiving a high-level prompt, while a human continues to supervise the process.
Full autonomy would be a further level in which the AI performs the work without a human in the loop. Anthropic said Claude has not reached that level in any of the R&D work covered by its measurement.
That distinction matters because the 26% figure does not mean Claude is independently designing and training its own successor.
Instead, it suggests that for a growing share of Anthropic’s research tasks, employees can assign Claude a broader objective and allow the system to work through much of the process while remaining subject to human oversight.
AI Is Involved in More Than 90% of Anthropic’s Research
Claude’s role extends beyond the 26% of work classified as AI-led.
Anthropic said more than 90% of its AI R&D is now performed at a level where AI at least “collaborates” with human researchers.
The company built what it calls the Anthropic R&D Automation Index to measure the trend. Researchers mapped thousands of individual R&D activities, grouped them into categories and assessed the degree to which Claude was involved in completing each type of work.
Anthropic said its analysis was based partly on roughly 15,000 granular R&D tasks collected from internal work records during July. The resulting framework includes hundreds of categories spanning areas such as model training, reinforcement learning, infrastructure, evaluations and serving systems.
The development reflects a wider move toward agentic AI, in which artificial intelligence systems can carry out longer, multi-step assignments rather than simply responding to individual prompts.
Business Innovative News has previously examined how this shift toward [agentic AI is changing business workflows in 2026].
Anthropic Says About 30,000 AI Agents Were Working Internally
Anthropic also disclosed the scale at which AI agents are already operating inside the company.
Approximately 30,000 agents were performing research and engineering work at any one time on Anthropic’s most widely used internal agent platform in August, according to the company.
Anthropic said every action taken by agents on that platform passes through an online monitoring system before execution.
After analyzing more than one billion agent decisions during August, the company said about 0.002% — roughly one in every 47,000 decisions — were blocked by the monitoring system.
Anthropic also operates post-action monitoring. Its system flags about 100,000 transcripts each week for further automated screening, with approximately 50 of the highest-priority cases escalated to human reviewers, according to the company.
The company said increased automation makes such oversight particularly important because agents are being given longer-running tasks and can increasingly delegate work to other agents.
Safety Receives Part of Anthropic’s AI Research Compute
Anthropic also published new data on how it allocates computing resources between AI capability development and safety research.
During a sample period from July 13 to July 20, about 6% of the computing power used for AI R&D was allocated to safety-related work, the company said.
Among research conducted by AI itself, the safety share was approximately 12%.
Anthropic described those estimates as conservative because work that contributed equally to safety and model capabilities was classified as capability research rather than safety research.
The company acknowledged that compute is an imperfect way to measure safety investment because some safety research can require significant researcher time without consuming the enormous computing resources associated with training frontier models.
Why Anthropic Is Publishing the Numbers
Anthropic said it intends to make measurements like these available regularly to give governments, independent researchers and the public greater visibility into what is happening inside frontier AI laboratories.
One question attracting growing attention is recursive self-improvement — the point at which an AI system could potentially play such a large role in developing its successor that the process requires increasingly little human involvement.
Anthropic said measuring AI-led R&D can help researchers assess how close the industry is to such a scenario.
The company also plans to give independent third-party evaluators access to internal processes, systems and data so they can verify safety practices, examine incidents and monitor key development metrics.
The disclosure comes as leading AI companies face growing scrutiny over how quickly advanced systems are progressing and how much control human researchers retain as AI agents become more capable.
Reuters reported that OpenAI has also moved toward publishing regular information about unexpected or unauthorized behavior by its AI systems.
Anthropic’s Growing Role in the AI Race
The disclosure arrives as Anthropic is attracting heightened attention from investors and the wider technology industry.
Business Innovative News recently reported that speculative pre-IPO trading has placed an implied valuation above $2 trillion on Anthropic, although that figure is not the company’s latest official valuation and reflects market expectations rather than a completed financing transaction.
The company’s latest automation data highlight another dimension of the AI race: increasingly powerful models are not only becoming commercial products but are also becoming tools used to create the technology that follows them.
For Anthropic, the 26% figure provides evidence of how quickly that transition is occurring — while the company’s own data also show that human oversight remains part of the process.
