Primary source AIArticle·18 September 2026·8 min read

Anthropic Lifts the Hood: Claude "Leads" 26% of Its AI R&D

For the first time, a frontier lab has published numbers on how fast AI is building its own successor — and on what it's doing to keep watch over it.

ShareWhatsApp X LinkedIn Facebook

Save to my selection

Anthropic Lifts the Hood: Claude "Leads" 26% of Its AI R&D

Source : Anthropic (Marina Favaro, Phillie Wright) · anthropic.comView original ↗

Who covers it

On TechmemeHacker News 5 pts ↗

Picked up by 13 outlets · 2 top-tier · 3 tech · 2 forums · 21 social posts

In brief

Anthropic proposes three families of metrics meant to make visible, from the outside, the pace of AI development inside labs: how much of its own R&D is automated, how its agents are supervised, and how much compute goes to safety. In-house snapshot: Claude "leads" 26% of Anthropic's AI R&D work, more than 90% of the work is at least at "collaboration" level, roughly 30,000 agents run internally, and 6% of R&D compute goes to safety. The company also announces the arrival of independent third-party evaluators embedded internally.

🍺 Bar-stool version

Anthropic just published the dashboard for its own kitchen: over there, AI already runs a solid quarter of the research work used to build the next AI. Don't worry, there's still a human in the loop — a bit like an airplane's autopilot, except nobody quite knows what altitude we're at. Along the way, we learn that 6% of compute goes to safety, which, presented as a selling point, says a lot about where the bar sits. Mostly, it's the first time a lab has agreed to put verifiable numbers on a question everyone's been asking for three years: exactly how fast is this going?

Key takeaways

1
Anthropic publishes an in-house "R&D Automation Index": as of August 2026, Claude "leads" (AL4 level) 26% of the company's AI R&D work, and more than 90% of the work is at least at "AI collaborates" level.
2
No measured subset reaches full autonomy (AL5): a human always remains in the loop, at least to validate and deploy.
3
Roughly 30,000 agents work simultaneously on the main internal research and engineering platform; 100% of their actions pass through an online monitor before execution.
4
Out of more than a billion agent decisions in August 2026, 0.002% were blocked (about 1 in 47,000); ~100,000 transcripts are flagged each week, of which ~50 are escalated to humans.
5
During the week of July 13–20, 2026, roughly 6% of AI R&D compute went to safety, and 12% of AI R&D compute was AI-driven.
6
Anthropic announces the integration of independent third-party evaluators, with access to internal processes and data comparable to that of internal risk-evaluation teams.
7
The automation scale used (AL0 to AL5) is the one proposed by Epoch AI, and Anthropic is pushing for it to become a shared methodology across labs.

The bet: measuring inputs, not just capabilities

Capability evaluations tell you what a model can do.

26%: what the number means (and doesn't)

The index relies on Epoch AI's scale, from AL0 (no AI) to AL5 (full autonomy).

30,000 agents and a control room

The second component is the most operationally concrete.

6% of compute for safety: the most debatable number

During the week of July 13–20, 2026, roughly 6% of the compute allocated to AI R&D went to safety work, and roughly 12% of AI-driven R&D compute did.

The real deliverable: a methodology others could copy

For each of the three measures, Anthropic specifies what "any frontier developer could publish today." This is the heart of the maneuver: turning an in-house report into a de facto standard, before a regulator invents a worse one.

Why it matters

This is the first time a frontier lab has put verifiable numbers on the question shaping the entire AI regulation debate: how fast are models participating in their own creation?

Older than 30 days: the archive

This article is more than 30 days old

A free account opens articles from the last 30 days in full. The full archive is in the Pro plan.

Create my account

Already a member? Log in

Compare plans →

#ai #anthropic #governance #agents #safety #compute

Original source

Measurements for understanding the pace of AI development inside frontier labs

Anthropic (Marina Favaro, Phillie Wright)

Open the article ↗

For you

Put it to work on your sources.

Free: a month of articles and three sources of your own. Pro: the whole archive and your sources, from €8/month.

Create my account →

For your team

The same machine, on your topics.

A space in your colours, your watch angles, your curators. Pilot open to three companies.

Request pilot access →