On September 17, Anthropic released a rare set of internal numbers: as of August 2026, Claude "leads" about 26% of the company's AI R&D work (under 1% in February), and more than 90% of R&D is at least at the "AI collaborates" level; at any moment roughly 30,000 AI agents run on the internal platform, making over 1 billion decisions in August alone. On the other side, OpenAI's official blog states it has met last fall's goal of an automated research intern and is progressing toward an automated AI researcher by March 2028 — while explicitly saying that fully autonomous RSI has not happened today and should not be pursued unless it can be done safely.
[1][2]The nut graf: RSI is moving from science fiction to a company-level KPI — when AI starts building the next generation of AI, the most notable thing is not whether it is happening but how publicly divided the labs are: Anthropic quantifies progress for you (26%, 30,000 agents, 1 billion decisions), OpenAI breaks the goal into milestones (research intern, then a 2028 researcher) with a safety conditional attached, and xAI's Musk says plainly that humans are increasingly out of the loop. Three statements, three answers to who holds the wheel.
On attribution first. Anthropic's 26% and 90% are self-reported internal statistics; the precise criteria for Lead versus Collaborate are not fully public; the 30,000 agents and 1 billion decisions are also company figures. OpenAI's statements come from its official blog (early September); the automated research intern and the March 2028 automated AI researcher are official goals rather than delivered commitments, and the blog explicitly conditions RSI on safe implementation ahead of speed. Read together, the substantive split is not technical route but narrative strategy: Anthropic uses data to argue AI building AI is already happening; OpenAI uses milestones to argue it has not happened yet and is in no rush. That divergence is itself the signal — with every frontier lab acknowledging the RSI direction, regulators and the public receive information that depends heavily on statistics each lab chooses to release.
Commercially, the RSI data makes research-automation rate a comparable company metric for the first time: how much AI each lab uses to build the next model feeds directly into efficiency narratives and valuation logic. Anthropic's 26% figure was translated into multiple languages within a week, which is the market treating it like an earnings disclosure. On risk, all three statements share the baseline that fully autonomous RSI does not yet exist — a common factual floor for the runaway-AI debate. But it is only a floor: whoever crosses the fully autonomous threshold first simultaneously gains the efficiency advantage and draws the most concentrated safety scrutiny, and that seat is still empty.
[1][2]