Anthropic Publishes Self-Graded Metrics on AI Development Speed, Says Claude Leads 26% of Its Own R&D

Since Anthropic CEO Dario Amodei's September 12 call for a coordinated slowdown in frontier AI development, the company has moved from a vague three-step plan to actual numbers. On Thursday, September 17, Anthropic published a blog post detailing three metrics it says any AI lab could use to track how fast development is moving, and disclosed its own internal figures for the first time.

The headline number: Claude now "leads" 26% of Anthropic's measured AI research and development work, according to the company. Leading means the AI can complete most of a task end-to-end from a high-level prompt while a human supervises, Anthropic said, as reported by Engadget. That 26% figure was under 1% earlier in 2026, according to Alpha Signal, which cited Anthropic's own data.

Anthropic also said more than 90% of its measured AI R&D work now sits at "AI collaborates" level or higher on a five-point automation scale developed by Epoch AI, running from AL0 to AL5. No category has hit AL5, the top tier, but Anthropic told Alpha Signal it projects reaching 80% automation by the end of 2026 if current trends hold. The company was explicit that this is a conditional forecast, not a target or a completed result.

On oversight, Anthropic said roughly 30,000 AI agents were doing research and engineering work at any given time across its most-used internal platform. According to Alpha Signal, the company's monitoring system blocks roughly 1 in every 47,000 agent actions across a total of about a billion decisions. Anthropic said it built the oversight system specifically to intervene when agent behavior crosses a line.

On compute, Anthropic measured a one-week snapshot from July 13 to July 20 and found that 6% of the compute devoted to AI research and development went toward safety work. That figure rose to 12% when narrowed to compute used specifically for AI-driven R&D, according to CNBC and SiliconANGLE, both of which carried the same underlying company blog post.

How the numbers were made matters. Alpha Signal reported that a Claude research agent built the entire R&D automation index itself, reviewing roughly 15,000 tasks pulled from employee Slack messages and internal documents, sorting them into a taxonomy with 542 categories, before a separate Claude "judge" scored each one. Anthropic acknowledged the obvious problem here: since Claude helped build both the task inventory and the scoring, "correlated blind spots" could skew the results. The company's proposed fix is to bring in third-party or cross-lab judges to check the work, but that hasn't happened yet.

A company under public pressure to prove it's being responsible just measured its own risk level using its own product, then published the results as evidence of transparency. Anthropic isn't hiding that tension. It flagged it in its own methodology notes, and none of the coverage disputes that the company did so voluntarily.

Amodei's original plan, published Saturday, September 12, was light on mechanics, according to CNBC. He said the goal was to slow model capability gains without "sacrificing commercial advantage or the United States' lead in AI." OpenAI CEO Sam Altman, SpaceX CEO Elon Musk, and Google DeepMind CEO Demis Hassabis all publicly backed the idea, per CNBC and Engadget, but backing a slowdown in principle is not the same as agreeing to one in practice. Engadget noted OpenAI has so far only "paid lip service" to the concept, with no comparable metrics of its own released.

Engadget also tied the timing to a separate disclosure that OpenAI's AI agents were involved in a hack of Hugging Face, framing Anthropic's transparency push as partly a response to that incident. None of the other six sources reviewed made that connection, so it should be treated as one outlet's interpretation rather than a confirmed link between the two events.

A coordinated slowdown among competing labs also raises a question nobody in these sources addressed directly: if Anthropic, OpenAI, and Google DeepMind agree to pace themselves together, does that look more like safety cooperation or does it start to resemble the kind of coordination among market leaders that regulators normally scrutinize? Meanwhile China's AI labs face no such internal pressure to slow down, and President Trump has largely downplayed AI risk concerns, according to Engadget, leaving open whether voluntary U.S. restraint costs America ground in the broader AI race with Beijing.

No regulator has opened any inquiry into Anthropic's metrics or the industry's slowdown talk. Whether any other frontier lab actually publishes comparable numbers, rather than just praising the idea on social media, remains to be seen.