How fast is AI accelerating its own development inside labs?

Anthropic argues that the public needs visibility into how fast frontier AI development is moving inside the labs building it, and it publishes a first snapshot of what that visibility could look like. The starting premise is that AI systems "have begun to automate more of the process of building themselves." The snapshot gives three figures. "Claude 'leads' 26% of Anthropic's AI R&D work," and the share of work "at or above 'AI collaborates'" is "above 90%." As of August 2026, "approximately 30,000 agents" were doing research and engineering work at any one time on Anthropic's most-used internal platform. The excerpt says those agent figures cover that platform only.

The reasoning is that labs increasingly use AI to build future models, which lets them "develop more capable models more quickly" and test them more before release, but that models accelerating their own development "could make it more challenging for humans to understand or control these systems." Measurement is meant to show "how close the world is to reaching recursive self improvement (a model fully autonomously building its successor)." The excerpt sets out three measurement areas: the share of R&D work along a path "from 'AI collaborates' toward 'AI leads'"; the oversight of semi-autonomous agents that delegate to one another; and compute allocation, which Anthropic calls "among the most verifiable inputs to the AI R&D process" and a possible "critical lever in a future pacing effort." Anthropic also says the numbers "would shift if there were coordination on pacing the frontier, as called for by Anthropic CEO Dario Amodei," and it plans to embed independent third-party evaluators with access comparable to internal risk assessment teams.

The nearest notes approach the same question from other angles. The Atria Dawn analysis reads task records to see who proposes and who decides, and Anthropic's scale asks a similar who-leads question at lab level, but as a self-reported share rather than task-level data. ASI-Bench's staged withdrawal of guidance tests how far an agent can go unaided on a benchmark, while Anthropic's scale tracks a comparable progression inside a working lab. The ASI-Evolve loop of analysis and reuse is the kind of AI-driven research cycle a lab-wide share would have to count, though the excerpt does not say whether such loops appear in the counted work. The efficiency bridge in the AI R&D benchmark note uses a different unit altogether: research efficiency under an evaluation budget, not a share of work.

The excerpt does not establish how "leads" or "collaborates" are assigned, whether shares are counted by time, tasks or tokens, or what the methodological Appendix says, because that Appendix is not in the excerpt. The 26% and above-90% figures are Anthropic's own internal numbers, and the independent evaluators it plans to embed are described as a future step. The excerpt gives one date for the agent count and no time series, so it cannot show whether the share is rising. The monitors that constrain the agents are "summarized below," but that summary is not in the excerpt, so how they constrain the agents is unstated. What follows is a self-reported snapshot from one developer on one platform. By the source's own account, it would need third-party verification and regular publication before it could show how the pace is changing.