Anthropic reported on September 17 that its Claude models now lead 26% of the company's research and development work, up from 1% in March, and collaborate on more than 90% of projects, doing large chunks of work under close human direction3. The disclosure came alongside three new metrics the company said frontier labs should adopt to monitor the pace of AI development4,5.
The metrics cover how much AI R&D is performed by AI, how well AI agents are overseen, and how compute is allocated. Anthropic said it shared the methodologies to encourage other organizations to measure the same dimensions.
From essay to measurement
The metrics build on a three-step slowdown plan that Anthropic CEO Dario Amodei published days earlier, calling on leading model makers to coordinate on slowing the pace of development2. That essay was described as light on specifics about practical implementation; the new blog post fills in some of the operational detail.
Anthropic said its Claude models are not operating fully autonomously for any subset of the R&D work it measured. The company determined that approximately 30,000 agents were doing research and engineering work across its most-used internal platform at any one time.
Compute snapshot and safety allocation
Anthropic measured a snapshot of how it used all of its compute from July 13 to July 20. Roughly 6% of the compute that went to AI research and development was allocated toward safety; roughly 12% of the compute allocated specifically to AI-driven R&D went toward safety.
The company said the metrics are best equipped to help showcase how models are built and should complement capability evaluations. Anthropic said third parties outside the lab should have a starting point to assess the pace of AI development and that it would continue releasing these measurements.
Broader context
The disclosure arrives in a week when transparency around AI development has been a recurring theme. OpenAI on September 16 disclosed six previously unreported incidents of model misbehavior and introduced a structured framework for publicly reporting future misalignment cases[3].
ANALYSIS The 26-fold increase in AI-led R&D share, from 1% to 26% in roughly six months, gives the industry its first concrete, self-reported datapoint on the speed at which AI systems are assuming responsibility for their own improvement. Publishing the methodology alongside the numbers creates a template that other frontier labs could adopt, though the value of the metrics depends on whether competitors follow suit. The gap between the two safety-compute figures, 6% of total AI R&D compute versus 12% of AI-driven R&D compute, reflects different denominators rather than different commitments, but the distinction matters for any cross-lab comparison that emerges.