Anthropic has opened an unusually detailed window into how its future AI models are being built. For the first time, the company has published a set of internal measurements showing how much research and engineering work Claude is already performing inside the lab — effectively measuring the extent to which an AI system is helping create the generations that come after it.
The headline figure is striking: as of August 2026, Claude “leads” 26% of Anthropic’s measured AI R&D work. Under the company’s definition, that means the model can take a high-level task, perform most of it end-to-end, deal with unexpected problems along the way and return the result for a human to review. Anthropic stresses that Claude is not yet fully autonomous in any measured category.
What “leading 26% of R&D” actually means
Anthropic uses an automation scale based on a framework developed by Epoch AI. At AL3, an AI system “collaborates”: it can complete large portions of a task but remains under close human direction. At AL4, it “leads”: a human can provide a high-level objective and let the agent work through most of the task independently before reviewing the outcome. AL5 would represent full autonomy, where the system can notice, scope, execute and deploy work without needing a person to initiate the process.
Claude has reached the AL4 threshold for roughly one quarter of Anthropic’s measured research and development. An even broader number shows how deeply AI has already entered the company’s workflow: more than 90% of Anthropic’s AI R&D is now performed at least in collaboration with Claude. Work with little or no AI involvement is becoming a minority of the measured development process.
The shift has been rapid. Anthropic’s published chart shows the share of work Claude “leads” rising from below 1% in February 2026 to 26% in August. Reuters described the disclosure as one of the clearest public windows yet into how frontier labs are using AI to develop the next generation of AI systems.
The key distinction: Anthropic is not claiming that Claude can independently design, train and deploy its own successor. The company is measuring an earlier stage, in which AI performs substantial portions of research and engineering while humans still set objectives, control infrastructure and approve consequential outcomes.
Roughly 30,000 agents are working concurrently inside Anthropic
The second major disclosure concerns scale. Anthropic says that in August, approximately 30,000 AI agents were doing research and engineering work at any one time on its most heavily used internal agent platform.
This is no longer a simple environment where individual employees open a chatbot for occasional assistance. Anthropic describes agents that can operate for extended periods and communicate with one another. Each agent is assigned its own persistent identity, and messages pass through a shared communication system so actions can be traced back to the specific agent that produced them.
That scale changes the safety problem. An unusual or unwanted action may be extremely rare for a single agent, but once tens of thousands of agents make enormous numbers of decisions, low-probability events can become operationally relevant.
Anthropic monitored more than a billion agent decisions in August
The company describes two layers of oversight. Online monitors examine actions before execution and can block or redirect them. Offline monitors analyze activity after the fact, looking for slower-developing patterns that would be difficult to identify from one isolated action.
Anthropic says 100% of the actions on the measured internal platform pass through its online monitor before execution. Across more than one billion decisions analyzed in August, approximately 0.002% — around one in 47,000 — were blocked. That figure should not be read as one malicious act in every 47,000 decisions; it represents actions the monitoring system considered risky enough to stop automatically.
The offline layer flags roughly one or two transcripts per thousand for additional review. About 100,000 transcripts per week go through further automated filtering, and Anthropic says the approximately 50 highest-priority cases each week are escalated to human reviewers.
The company also disclosed how much compute goes to safety
Anthropic’s third metric looks at computing resources. The company analyzed one week, July 13 through July 20, and classified AI R&D workloads according to whether their primary purpose was advancing capabilities or improving safety.
It estimates that about 6% of AI R&D compute during that week went to safety work. Among research tasks performed by AI systems themselves, the share was about 12%. Anthropic calls the methodology conservative because workloads that advanced safety and capabilities equally were counted as capability work rather than safety work.
The company also cautions against treating compute share as a universal safety score. Some safety research consumes relatively little accelerator time but a great deal of human effort. The metric is therefore more useful for observing changes over time than for declaring one laboratory safer than another.
Why Anthropic wants these numbers published regularly
The broader purpose is to create a measurable indicator for how close frontier labs may be getting to AI systems that substantially accelerate the development of future AI. Anthropic uses the term “recursive self-improvement” for the limiting case in which a model could autonomously build its successor. It explicitly says Claude has not reached that point.
What is new is the attempt to replace vague claims about “AI building AI” with numbers that can be tracked over time. Anthropic argues that frontier developers could regularly publish the share of R&D conducted by AI, the effectiveness and coverage of agent monitoring, and the distribution of compute between capabilities and safety.
The company is encouraging other labs to report comparable metrics, although it acknowledges that meaningful cross-company comparisons are not yet straightforward. There is no universal methodology, and laboratories may draw different boundaries around what counts as collaboration, leadership or autonomous work.
An important limitation: Claude is partly evaluating Claude
The technical appendix contains perhaps the most important methodological caveat. Anthropic constructed a dataset of roughly 15,000 granular R&D tasks from internal work records. Claude helped organize those tasks into a hierarchy containing 542 nodes. Claude agents then investigated how each kind of task was performed, and a separate Claude judge assigned the automation level.
Anthropic did compare the model’s judgments with human employees. Exact agreement between model and human ratings was 59%, while human-to-human exact agreement was 35%; model and human ratings were within one automation level 97% of the time. Even so, the company acknowledges the obvious concern: an AI system is involved in measuring how much work that same AI system is doing.
That is one reason Anthropic says independent verification will matter. The company plans to embed outside evaluators from multiple organizations and give them access to internal processes, systems and data comparable to what its internal risk teams can inspect.
What has actually changed
The disclosure does not mean Claude has suddenly begun autonomously designing its own replacement. The more consequential shift is subtler: frontier AI development is moving from a workflow where people use AI tools toward one where large fleets of AI agents are becoming direct participants in research and engineering.
If the share of AL4 work continues rising, model intelligence will not be the only variable that matters. Monitoring coverage, infrastructure permissions, human approval boundaries, independent auditing and the ability to attribute actions to individual agents will become increasingly important parts of the development stack.
For now, Anthropic’s numbers provide a rare quantitative marker: Claude can lead 26% of measured AI R&D nearly end-to-end under supervision, participates substantially in more than 90%, and operates as part of a system involving roughly 30,000 concurrent agents — while the company still reports no fully autonomous AI R&D category.
