New Anthropic AI development metrics reveal Claude AI R&D automation at 26%, 30,000 agents running internally, and only 6% of compute going to safety.
Anthropic published three new measurements on Thursday that track how fast AI development moves inside frontier labs. The company released its own numbers first and shared the methodology so rivals can match it. The move follows CEO Dario Amodei’s call for a coordinated industry slowdown six days earlier.
Claude AI R&D Automation Climbed From Under 1% to 26%
Anthropic’s new R&D Automation Index rates Claude as “leading,” at 26% of the company’s AI research work as of August 2026. At that level, Claude finishes a task end-to-end from a high-level prompt while a human reviews the output.
More than 90% of the work now sits at or above the level where AI collaborates. In February, the leading share sat below 1%. No category has reached full autonomy.
Anthropic graded the work using an automation scale from Epoch AI, running from no AI involvement up to fully autonomous.
Claude Measured Claude’s Own Automation
The methodology carries the real story.
Anthropic sampled 20% of staff per department each week of July and sent a Claude agent through their Slack messages and internal docs to reconstruct what each person did. That pass yielded roughly 15,000 tasks. Claude sorted them into 542 categories. A second Claude agent researched each one. An independent Claude judge assigned the grade.
The company admits the judge can inherit the blind spots of the model it grades. Its agreement study cuts both ways: the model matched human raters exactly 59% of the time, while humans matched each other just 35% of the time.
Anthropic has graded its own technology in public before, most recently with its observed-exposure study on AI job displacement.
AI Agent Oversight: One Block in Every 47,000 Decisions
Roughly 30,000 agents run research and engineering work at Anthropic at any moment on its main internal platform. Online monitors screen every action before execution and block irreversible moves such as an agent copying model weights out. Offline monitors ingest everything afterward and hunt for slow drift.
Anthropic reviewed more than a billion agent decisions in August and blocked 0.002% of them, about one in 47,000. The offline system flags around 100,000 transcripts weekly and escalates roughly 50 to humans.
The company issues its own warning: agents rarely misbehave individually, but rare events turn routine once millions operate at scale. That gap between stated risk and shipped product has drawn scrutiny before.
Each agent also carries a persistent identity that survives model upgrades, and agents talk in a shared open channel so monitors can trace how a conclusion spread.
AI Safety Compute Sits at 6%, and Anthropic Says the Number Misleads
Over one week in July, 6% of AI R&D compute went to safety work. Within AI-driven R&D specifically, safety took 12%.
Then Anthropic undercuts its own scorecard. Compute share measures spending, not effort. A faster safety classifier lowers the percentage without reducing the safety work. The company reports exactly that effect in its own data.
Anthropic publishes capability results separately through its Responsible Scaling Policy framework, which shifted to a tiered ASL-3 structure this year.
No Rival Lab Has Signed On
Anthropic frames the three measurements as a proposal, not an agreement. OpenAI, Google DeepMind and others have committed to nothing. Verification remains a plan: the company says it will embed independent evaluators with internal-grade access.
The same restraint logic drove its decision to hold Claude Mythos back from public release.
What exists today is a self-report, produced by the system it describes, scored against a rubric the company wrote. The load-bearing claim is not that Claude leads 26% of AI R&D. It is that someone outside the building could eventually check.
Related: AI co-workers reshaping office work
