Anthropic: Claude Now Leads 26% of Its Own AI R&D

Anthropic's first R&D Automation Index reveals Claude 'leads' 26% of internal AI research, with oversight monitors covering 30,000 agents.

@sharedot

Anthropic: Claude Now Leads 26% of Its Own AI R&D

By @sharedot · · 7 pages

  • AI Frontier
  • AI Safety
  • Frontier Labs
  • Agents

Anthropic's first R&D Automation Index reveals Claude 'leads' 26% of internal AI research, with oversight monitors covering 30,000 agents.

What Anthropic just disclosed

The centerpiece is the Anthropic R&D Automation Index, built by cataloguing every kind of AI research work at the company, rating how automated each task is on Epoch AI's AL0–AL5 scale, and aggregating the results.

As of August 2026, Claude 'leads' 26% of Anthropic's AI R&D — completing most tasks end-to-end from a high-level prompt under human supervision — and more than 90% of work sits at 'AI collaborates' or above. Claude operates fully autonomously on no measured subset yet.

Why the number is surprising

Labs have long described AI-assisted research in vague terms; here Anthropic concedes that a quarter of the work designing Claude's successors is already led by Claude itself.

Anthropic frames the disclosure as the raw material for public pacing decisions, noting CEO Dario Amodei has called for coordinated slowing of the frontier and that these figures would shift if such coordination happened.

It is a reversal of the usual silence: the lab is handing outside parties the yardstick.

The oversight evidence inside the lab

Anthropic reports roughly 30,000 agents doing research and engineering work on its most-used internal platform at any moment in August 2026, all on its own systems. Every agent action passes an online monitor before execution, and 100% of actions are ingested offline afterwards.

Over a billion agent decisions analyzed that month saw just 0.002% — about 1 in 47,000 — blocked, with roughly 100,000 transcripts flagged weekly and only about 50 escalated to humans. Anthropic says individual agents rarely misbehave in its data, but warns rare events compound at economic scale.

Compute and the safety share

The third metric slices Anthropic's compute.

Over the week of July 13–20, about 6% of compute going to AI R&D was allocated to safety work, and about 12% of compute going to AI-driven AI R&D was safety — figures Anthropic calls deliberately conservative, excluding safeguard classifiers, which represent a separate, comparable amount of compute.

Anthropic concedes compute is an imperfect proxy for safety focus, since safety research is researcher-time-heavy but compute-light; the value is a like-for-like comparison across labs and over time, with third parties checking the classification.

The stakes and what comes next

The disclosures land amid a real coordination push.

According to Bloomberg via Stocktwits, OpenAI policy chief Chris Lehane said OpenAI has spent weeks in safety discussions with Anthropic and Google DeepMind and backs bipartisan federal AI risk legislation, while Sam Altman has endorsed Amodei's proposal for independent evaluators with permanent, employee-level lab access.

European Commission President Ursula von der Leyen told the European Parliament she will invite the main frontier labs to discuss pacing, citing cyberattack risks from self-recursive models.

Sources

  1. anthropic.com › Measurements for understanding the pace of AI development inside frontier labs
  2. stocktwits.com › OpenAI, Anthropic And Google DeepMind Discuss AI Safety As Industry Faces Pressure To Slow Frontier Development: Report
  3. businessworld.in › EU Chief Von Der Leyen Calls For Slower Pace On Frontier AI

More on AI Frontier