Claude leads roughly a quarter of the AI research and development happening inside Anthropic, and more than 90% of that work now involves the model as at least a collaborator. That’s the headline figure from a new report by Anthropic, which published three internal measurements meant to give outsiders a way to track how fast frontier labs are moving. Nothing here is fully autonomous yet. But the numbers show how much of the work of building the next model is already being handed to the current one.
What stands out here is the framing. Anthropic isn’t reporting this as a capability brag. The company says the goal is to shrink the gap between what labs know and what the public knows, as the industry debates whether to slow down frontier development.
The three measurements
Anthropic proposes tracking three things, and shares a snapshot of each from inside the company:
- AI-led AI R&D: how much of the work of building models is done by AI itself
- Agent oversight: how well the lab can watch and intervene in what its agents do
- Compute allocation: what share of compute goes to safety versus everything else
How they measured automation
Anthropic built what it calls the R&D Automation Index. For every week in July 2026, it randomly sampled 20% of staff in departments involved in model R&D. A Claude research agent reviewed each person’s week through Slack and internal docs and listed their tasks. That produced about 15,000 granular tasks, which Claude then organized into a tree of 542 categories.
A separate Claude judge rated each category on a scale from Epoch AI that runs from AL0 (no AI involvement) to AL5 (fully autonomous). The two levels that matter most: AL3 means AI “collaborates,” doing big chunks of work under close human direction. AL4 means AI “leads,” completing most of a task end-to-end from a high-level prompt while a human supervises. Categories were weighted by how much person-time they consumed.
The results as of August 2026:
- Claude leads (AL4 or above): 26% of AI R&D work
- Claude collaborates or higher (AL3+): above 90%
- Fully autonomous (AL5): 0%
Anthropic checked the judge model against human raters who own the relevant work. Model-versus-human exact agreement was 59%, while humans only agreed with each other 35% of the time. Model and human ratings landed within one level of each other 97% of the time.
Oversight: 30,000 agents, one in 47,000 blocked
Anthropic reports that around 30,000 agents are doing research and engineering work on its main internal platform at any moment. Every action passes through an online monitor before execution, usually within seconds, and every action gets ingested by an offline monitor afterward.
Out of more than a billion agent decisions in August 2026, the online monitor blocked 0.002%, or about 1 in 47,000. The offline monitor flags one to two transcripts per thousand, roughly 100,000 per week. Classifiers filter those down, and the top 50 or so per week reach human review.
Two design choices make this work, according to Anthropic. Each agent gets a persistent identity that survives model upgrades, so its record is auditable over time. And agents communicate through an open, shared messaging system rather than privately, so they can check each other’s claims and monitors can trace conversations between them.
Compute: 6% to safety
Over one week in July, about 6% of Anthropic’s AI R&D compute went to safety work. Within compute used specifically for AI-driven AI R&D, the safety share was around 12%. Anthropic calls these deliberately conservative numbers. Anything that helped capabilities as much as safety was counted as R&D, and safeguard classifiers were excluded entirely.
The limitations Anthropic admits
The report is unusually candid about what it doesn’t prove:
- One week of compute data shows the measurement is possible, not that a trend exists
- Workload labels are best-effort and unverified
- Compute share measures spending, not safety output. A more efficient safety classifier lowers the number without meaning less safety work
- Claude is judging Claude, so the judge could share blind spots with the systems it evaluates
- The task basket is frozen to July 2026, so it can’t yet show whether humans are shifting to new kinds of work
Anthropic says it plans to embed third-party evaluators from multiple organizations with access comparable to its internal risk teams.
What you can take from this
If you run AI agents in your own work, the oversight section is a practical blueprint. Give agents persistent identities. Log every action. Put a fast automated check in front of irreversible actions and a slower review behind everything else. Track coverage, latency, and escalation rate. Anthropic runs this at 30,000 agents, but the same three metrics work at 30.
The bigger signal is that 26% number. Anthropic is asking every frontier lab to publish the same figure on a shared methodology. Whether or not competitors follow, it’s the first hard data point on how close we are to recursive self-improvement, and it’s a number worth watching each time it updates. Full methodology and the classifier prompts are in the appendix of the original report.