When AI builds itself - By Anthropic

TL;DR

• 1

  Claude now writes most of Anthropic's production code: As of May 2026, more than 80% of code merged into Anthropic's production codebase was attributed to Claude, up from low single digits before Claude Code launched in February 2025.

• 2

  Engineer output jumped sharply once agents could run code themselves: Anthropic says code merged per engineer stayed flat from 2021 through 2024, then rose in 2025 and again in 2026, reaching about 8x the 2024 level in Q2 2026.

• 3

  AI task horizons are stretching fast: Anthropic cites METR data showing the length of tasks models can complete reliably has been doubling about every 4 months, from roughly 4-minute software tasks in March 2024 to 12-hour tasks by Claude Opus 4.6 a year later.

• 4

  Claude is moving from assistant to researcher on bounded problems: On an internal model-optimization task, Anthropic says Claude improved from about 3x speedup in May 2025 to about 52x in April 2026, while a skilled human would need 4 to 8 hours to reach about 4x on the same setup.

• 5

  Judgment is still the key human moat, but Anthropic thinks it is shrinking: In internal research-session tests, Claude's suggested next step beat the human researcher's actual choice 64% of the time in April 2026, up from 51% in November 2025, though humans still set goals and see the bigger picture.

• 6

  Anthropic frames recursive self-improvement as plausible, not inevitable: The piece sketches three futures, from stalled progress to AI labs getting compounding efficiency gains to full recursive self-improvement, and argues a credible pause would require verification systems that do not yet exist.