Anthropic says Claude leads 26% of its AI research work

Anthropic published internal pace metrics for how AI is building AI inside the lab: Claude “leads” 26% of Anthropic’s AI R&D work (Automation Level 4), more than 90% of work sits at “collaborates” or above, about 30,000 research and engineering agents run on its main internal platform at once, and roughly 6% of AI R&D compute in a one-week snapshot went to safety.

As Washington and rival labs argue over pacing the frontier, Anthropic is putting numbers on recursive self-improvement inputs — automation share, agent oversight, and safety compute — and saying third-party evaluators should verify them.

The rule in English

Three measurements (Anthropic Institute post on measuring the pace of AI development):

1. AI-led AI R&D (R&D Automation Index). Epoch AI–style Automation Levels AL0–AL5. As of August 2026: Claude is not fully autonomous (AL5) on any measured subset; Claude leads (AL4) 26% of Anthropic AI R&D; share at or above AI collaborates (AL3+) is above 90%.

2. Oversight of AI agents. On the most-used internal platform: ~30,000 agents doing research/engineering at any one time (Aug 2026). Online monitors: 100% of actions checked before execution; ~0.002% blocked (~1 in 47,000) across >1 billion decisions in August. Offline monitors: 100% ingestion after the fact; ~100,000 transcripts flagged/week; ~50/week escalated to human review.

3. Compute allocation. Week of July 13–20, 2026: about 6% of compute going to AI R&D counted as safety; about 12% of compute going to AI-driven AI R&D counted as safety. Conservative definition (dual-purpose work counted as R&D, not safety); excludes separate safeguard-classifier compute.

Adjacent Institute piece (“When AI builds itself”). Public and internal evidence that AI is already accelerating AI development (e.g., engineers shipping far more code with Claude). Full recursive self-improvement — a system autonomously designing its successor — is not claimed as present. Anthropic says a verifiable multi-lab slowdown/pause would be good if others matched; unilateral pause is weaker.

Evaluators. Anthropic reiterates plans to embed independent third-party evaluators with internal-comparable access to verify practices, incidents, and metrics like these.

Who wrote it / who it binds

Writers: Anthropic Institute / company researchers (Favaro, Wright, Clark et al. credited on the measurements post). Binds Anthropic’s published transparency posture. Does not bind other labs or create a federal reporting duty.

Who it exempts

Competitors. Agent metrics cover one internal platform only. Compute share is one week, not a fixed budget line.

What a founder, parent, or worker should do differently

Founders at frontier labs: treat AL4 share, agent coverage/block rates, and safety-compute % as the comparison set Anthropic just published — match, beat, or explain why not.

Policymakers: these are the “verifiable inputs” Anthropic wants pacing debates to use; ask other labs for the same three numbers under third-party check.

Parents and workers: this is lab process transparency, not a consumer-product safety score.

What’s next

• Whether OpenAI, Google DeepMind, or others publish comparable AL / agent / safety-compute figures.
• Third-party evaluator access actually seating and re-running these metrics.
• Movement from AL4 lead share toward AL5 — the company’s own red line for full autonomy on R&D subsets.
• Whether governments turn these measurements into disclosure mandates.

Sources: Anthropic Institute, measuring pace of AI development; Anthropic Institute, “When AI builds itself”.