Anthropic outlines metrics to track AI development at frontier labs: how much AI R&D is done by AI, how well agents are overseen, and how compute is allocated
The disclosure also arrives after Anthropic data identified software engineering as roughly half of its agent tool-call activity, making AI’s role in the work that produces new models a material governance question. Public reaction focused both on the reported 26% Claude-led R&D share and on how broadly “AI R&D work” is defined.
First-order effects
Anthropic gives frontier labs three explicit reporting dimensions—AI-led R&D work, agent oversight, and compute allocation—and supplies a disclosed baseline for Claude’s role in its own R&D.
The framework makes the supervision of agents and the division of compute between development activities management metrics rather than solely internal operational choices.
Second-order effects
Other frontier labs face pressure to state whether they can produce comparable measures, while differences in how labs define AI-led work may limit direct comparisons.
Policymakers and external evaluators gain a more concrete basis for asking how rapidly labs are automating model development and what controls accompany that shift.
Third-order effects
If labs converge on common definitions, frontier AI governance can move toward recurring operational disclosure rather than broad capability claims, with compute use and agent control becoming auditable institutional signals.
Absent shared scope definitions, competing self-reported metrics risk becoming a legitimacy exercise rather than a reliable benchmark for comparing frontier labs.
The trend: Frontier AI labs are beginning to institutionalize measurement of AI-assisted research itself, linking development speed to oversight and compute governance.
Anthropic reports that Claude went from ‘leading’ 0% of AI model R&D work in February '26 to 26% in August '26. At that rate, Claude will lead 100% of R&D work by Feb. '28. RSI is very close. Things are accelerating much faster than the most bullish thought.
for the first time ever, Anthropic opened the doors and showed how fast AI is actually building itself and it's a lot: AI already runs 26% of the work that builds their next models, up from under 1% in February and could hit 80% by the end of the year. around 30,000 AI agents are…
I don't think this really answers the question of what is happening in “RSI”. Says Claude “leads” 26% of Anthropic's AI R&D work, but doesn't define what falls in scope of AI R&D work. I'm not satisfied by this, but it's a step in the right transparency direction.
Claude is already helping Anthropic build the next Claude. The share of model R&D led by Claude went from 1% to 26% in just five months. If a data pipeline breaks, for example, a human hands it the task. Claude checks the logs, finds the cause, writes a fix, runs tests, and keeps…
AI systems are getting more powerful, and they're increasingly being used to build the next version of themselves. We want to illuminate that progress for the public. Today, we're sharing three measurements that help track AI development: 1. How much AI R&D is done by AI. 2. How …
Anthropic has added another ad-hoc contribution to Pacing. We should recognize the reality that we are already pacing, just poorly and unstructured-ly! I think we can do better! And to get at that, let me talk about something near and dear to me as a former NYC quant trader. (I p…