FULL STORY

Anthropic Opens Up: Claude Now Drives 26% of Its R&D

Anthropic first revealed that ~80% of its code is written by Claude, then published three internal metrics showing Claude-driven R&D jumping from under 1% to 26% in six months.

2026-09-13 ~ 2026-09-21 · 4 episodes · 47 posts

Episode 1 · Anthropic: Claude Now Writes 80% of Its Engineers' Code (2026-09-13, 2 posts)

Anthropic reveals its engineers now ship 8x more code per quarter than the 2021-2025 baseline, with about 80% written by Claude and CI tasks up 25x in six months. The company says AI is already accelerating the development of AI itself, signaling approaching recursive self-improvement.

Episode 2 · Anthropic Discloses Internal Metrics: Claude Now Drives 26% of Its R&D (2026-09-18, 41 posts)

On September 18, Anthropic published three internal metrics on its official blog, giving the public a first systematic look at how "AI is being used to build next-generation AI" — a first among frontier labs. Internal snapshots show Claude now drives about 26% of Anthropic's R&D, with roughly 30,000 agents running concurrently.

Confirmed

  • Anthropic released three metrics: share of AI R&D done by AI, supervisability of AI agents, and compute allocation; it stated any frontier developer could publish the same metrics and that they can be verified by third parties
  • The work comes from Anthropic's economic research series on measuring internal AI development speed, and uses Epoch's six-level automation scale: Claude "leading" corresponds to AL4 — end-to-end completion of most tasks under high-level prompts, still under human supervision; the share rose from about 1% six months ago to about 26% as of August, with over 90% of R&D work at "AI collaboration" (AL3) level or above; no subtask has reached full L5 autonomy
  • Posts differ slightly on the starting point: under 1% in February (@Polymarket, @DrSingularity), about 1% in March (@markk), or zero seven months ago (@AxomaticallyExtinct)
  • About 30,000 agents run concurrently on Anthropic's most-used internal platform for research and engineering
  • Per @gaganghotra, only 6% of AI R&D compute goes to safety
  • Per @chrismattmann citing the official blog, since May about 80% of code written for Claude products was actually generated by Claude itself
  • Per @erikbryn, a Stanford-related panel is proposed as a third-party auditor for external validation
  • The Washington Post, AP, and Business Standard covered the disclosure, all citing the 26% figure; @TansuYegen argued it signals the trend better than another benchmark score; Polymarket opened a prediction market on it

Unconfirmed

  • User shawnup questioned the wording motive behind the "labs in democratic countries" qualifier, a personal view without Anthropic response
  • Commenter AryHHAry noted the work is narrower than reproducible science, covering only speed, supervision, and compute

Why it matters

  • This is the first time a frontier lab has quantified AI participation in its own R&D, providing a trackable public baseline for assessing AI self-improvement, though reproducibility is questioned
  • The jump from about 1% to 26%, combined with 30,000 concurrent agents and about 80% of product code AI-generated since May, suggests AI development speed may decouple from the human-research paradigm — a key datapoint for assessing intelligence-explosion risk; @markk notes the feedback loop has formed even without full autonomy
  • The 6% safety compute share has also fueled debate over whether safety investment matches acceleration

Timeline

  • February (or March): Claude-led internal AI R&D share under 1% (or about 1%)
  • From May: about 80% of code for Claude products generated by Claude
  • August: share reaches about 26%; over 90% of work at AI-collaboration level
  • September 18: Anthropic publishes the three metrics and internal snapshots

21 more related posts →

Episode 3 · Anthropic Unveils Three Public Metrics to Track AI Progress (2026-09-18, 2 posts)

Anthropic has released three publicly trackable metrics to measure AI progress, including the share of AI R&D work performed by AI systems, aiming to give the public visibility into frontier labs' internal pace of AI development.

Episode 4 · Claude Now Drives 26% of Anthropic's Internal AI R&D, Up From 1% (2026-09-20, 2 posts)

Anthropic disclosed that Claude's share of internally-led AI R&D tasks jumped from under 1% in February to 26% within six months, a milestone cited as evidence of accelerating recursive self-improvement.