Anthropic Opens Its Books: AI Now 'Leads' 26% of Its R&D

Anthropic published first-of-their-kind internal measurements showing Claude leads 26% of its AI R&D while labs hold cross-company safety talks.

Anthropic has published a set of measurement tools intended to give the public, third parties, and governments better visibility into the pace of frontier AI development. The company says its systems are becoming exponentially more powerful and have begun to automate more of the process of building themselves, and that the public needs more information as the world considers slowing the frontier. The post lays out three measurement areas: how much AI is building the next version of itself, the ability to oversee and intervene in AI agent actions on Anthropic's systems, and the resources powering more capable models. It pairs the methodology with a snapshot of current numbers from inside the company.

The centerpiece is the Anthropic R&D Automation Index, a prototype index cataloguing every kind of AI R&D work at the company, rating how automated each task is, and aggregating the results. Using Epoch AI's Automation Level scale — from AL0 (no AI involvement) to AL5 (fully autonomous, no human in the loop) — Anthropic reports that as of August 2026 Claude is not operating fully autonomously on any measured subset of work. But Claude 'leads' 26% of Anthropic's AI R&D work, meaning it completes most tasks end-to-end from a high-level prompt while a human supervises, and the share of work at or above the 'AI collaborates' level is above 90%. The company notes this falls short of recursive self-improvement, where a model fully autonomously builds its successor.

Anthropic reports that roughly 30,000 agents were doing research and engineering work at the company at any one time in August 2026 on its most-used internal platform. All of those agents' actions pass through an online monitor before execution, usually within seconds, and 100% of actions are ingested by an offline monitor after the fact. Across more than a billion agent decisions analyzed in August, 0.002% — about 1 in 47,000 — were blocked, and roughly one to two transcripts per thousand are flagged for further review. Humans review blocked actions within one week, and around 50 high-priority flags per week are escalated to human review. Anthropic notes METR has independently red-teamed its offline monitoring platform, and that it plans to embed independent third-party evaluators with access comparable to internal risk teams.

On compute, Anthropic examined a snapshot of all its compute use from July 13 to July 20 and sorted every workload into categories. It found about 6% of compute going to AI R&D was allocated toward safety work, and about 12% of compute going to AI-driven AI R&D was allocated toward safety — figures the company describes as deliberately conservative. Anthropic cautions that compute is an imperfect proxy for safety focus, since safety research is often time-consuming for researchers but not compute-intensive, and the metrics exclude safeguards classifiers, which it calls a separate, comparable amount of compute. The company argues the value is providing a straightforward way to compare like with like across developers and over time.

The disclosure lands amid wider industry movement. According to Bloomberg, as reported by Stocktwits, OpenAI global policy chief Chris Lehane said OpenAI has been in safety discussions for several weeks with Anthropic and Alphabet's Google DeepMind, and that OpenAI does not believe an antitrust waiver is necessary for those talks. The discussions focus on safety cooperation rather than a broader agreement to slow development. Dario Amodei recently called on the industry to 'pace the frontier' and proposed giving independent evaluators permanent, employee-level access to frontier labs, a proposal Stocktwits reports OpenAI CEO Sam Altman has backed. Stocktwits also reports OpenAI supports the bipartisan Obernolte-Trahan federal AI governance framework, while Senate legislation would address catastrophic risks and a Jim Banks–Adam Schiff proposal would create a narrow antitrust carveout for sharing loss-of-control and bio/cyber threat information.