On September 17, 2026, the Anthropic Institute published its inaugural R&D Automation Index, a prototype measurement of how much of the company’s AI research and development is performed by Claude. The disclosure arrived five days after CEO Dario Amodei published We Must Pace the Frontier, a 3,800-word argument for deliberate frontier deceleration that triggered immediate endorsements from Sam Altman, Elon Musk, and Demis Hassabis. The entity most vocal about the risks of frontier AI is simultaneously the entity most aggressively automating its own development with that same technology.
The index measures automation on a six-level scale developed by Epoch AI, ranging from AL0 (no AI involvement) to AL5 (fully autonomous, no human in the loop). As of August 2026, Claude “leads” – completing most of a task end-to-end from a high-level prompt under human supervision – 26% of Anthropic’s AI R&D work. That figure was less than 1% in February. Over 90% of the company’s R&D now involves some form of Claude collaboration at AL1 or above. Fully autonomous R&D (AL5) remains at zero. Anthropic’s new wet lab in the San Francisco Bay Area, where Claude directs robotic lab equipment for biological experiments, is one concrete data point in this acceleration.
These numbers come with a structural caveat: they are self-reported, and the index is a prototype. Anthropic has not subjected these metrics to independent audit, though the company plans to embed third-party evaluators with access comparable to internal risk assessment teams. For an industry where the gap between public safety rhetoric and operational reality is widening, the lack of external verification matters. The data is a proprietary snapshot, not an industry benchmark.
The timing of the disclosure is difficult to separate from the broader convergence of events defining the industry’s current instability. Amodei’s September 12 essay framed deceleration as a moral imperative. Five days later, the R&D Automation Index showed Anthropic accelerating. Six days after the index, Reuters reported that Anthropic is considering releasing a new model to counter GPT-6 Astra ahead of its expected IPO – contradicting the pacing thesis the company had just published. This follows OpenAI’s September 15 decision to rule out a 2026 IPO, citing safety obligations, and a September 18 antitrust class action alleging that Anthropic, OpenAI, SpaceXAI, and Google are operating as an output-restricting cartel under the guise of safety.
The structural question is straightforward: can the entity advocating for deceleration be trusted to decelerate when its own R&D engine is accelerating at this rate? The gap between the public-facing safety thesis – which Anthropic has built institutional infrastructure around, including the proposed FINRA-style safety body – and the internal operational reality suggests that safety may be functioning more as institutional positioning than as a binding constraint on capability development.
Anthropic’s own data makes the contradiction legible. The company reports that approximately 30,000 agents are doing research and engineering work at Anthropic at any given time. Online monitors block 0.002% of agent decisions – about 1 in 47,000. Six percent of compute allocated to AI R&D goes toward safety; 12% of AI-driven AI R&D compute is safety-focused. The company frames these measurements as transparency tools for a world considering pacing the frontier. But the measurements themselves describe a system where AI is rapidly automating the development of more capable AI, with safety oversight occupying a small share of the computational budget.
The contradiction extends beyond Anthropic. The evaluation tax that builders absorb with each frontier release – benchmarking, migration testing, safety review, procurement negotiation – compounds with every model cycle. When Anthropic, OpenAI, and Google ship models within the same week, the builder-side cost of keeping up is the operational reality of an industry that is not, in practice, slowing down. The safety thesis, whatever its intellectual merits, has not altered the cadence of capability development. If anything, the entities most invested in the safety narrative are the ones most aggressively automating their own R&D.