Full Breakdown
Anthropic Reports Claude Leads 26% of Its AI R&D Work
By Drooid · · How we work
Core Findings: Claude’s Expanding Role
On September 17, 2026, Anthropic PBC released its R&D Automation Index, showing Claude “leads” 26 % of the company’s AI research and development work as of August 2026. The index also indicates AI “collaborates” on more than 90 % of measured tasks, with Claude never operating fully autonomously in the evaluated categories.
Measurement Methodology
The index uses Epoch AI’s automation scale (AL0–AL5). Tasks are classified by a Claude-driven “judge” after a Claude agent extracts weekly work records from Slack and internal documents for a random 20 % staff sample per department. The judge’s agreement with staff ratings was 59 %, versus 35 % agreement among staff themselves; ratings fell within one level 97 % of the time. The methodology is limited to a frozen basket of existing tasks and does not capture newly emerging work types.
Scale of Agent Activity and Oversight
Anthropic reported roughly 30,000 AI agents active on its primary internal platform in August 2026. Over the month, more than one billion agent decisions were logged. An online monitor screened every action, blocking 0.002 % (about one in 47,000) and forwarding those cases to human reviewers within a week. Offline monitors later ingested all actions, flagging roughly 100,000 transcripts per week; about 50 high-priority flags received human review.
Compute Allocation to Safety
A snapshot from July 13–20, 2026 showed 6 % of compute devoted to AI R&D allocated to safety work, rising to 12 % for AI-driven AI R&D. Anthropic describes these figures as conservative, noting that work advancing both capability and safety was counted as capability work and that separate safeguard classifiers were excluded.
Official Statements & Responses
Criticism & Opposition
- Researcher Jacob Coxon resigned, accusing AI firms of “gambling with our lives” and warning of existential threats from accelerating model capabilities.
- Critics note the 26 % figure is self-reported and not independently verified, questioning a methodology that relies on the same AI system it evaluates.
Verbatim Quotes
- “Models accelerating their own development could make it more challenging for humans to understand or control these systems,” — Anthropic
- “The share of work at or above 'AI collaborates' is above 90%,” — Anthropic
What’s Next
Anthropic plans regular publication of similar measurements and will embed independent third-party evaluators with access to internal processes, systems, and data. The company hopes a shared public methodology will enable cross-lab comparisons and support future transparency obligations.