---
id: "2026-06-04-anthropic-when-ai-builds-itself"
url: "https://postcutoff.com/e/2026-06-04-anthropic-when-ai-builds-itself/"
as_of: "2026-10-09T19:24:00+02:00"
date: "2026-06-04"
date_precision: day
category: policy-safety
importance: 4
confidence: high
status: [Confirmed]
sources: 4
editor: Adam Bicz
human_review: null
version: null
---

As of: 2026-10-09 19:24 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/e/2026-06-04-anthropic-when-ai-builds-itself/

# Anthropic's 'When AI builds itself'

Full title: Anthropic's 'When AI builds itself': Claude writes 80%+ of merged code, engineers ship 8x more; Anthropic would join a verifiable multi-lab pause

On June 4, 2026 the Anthropic Institute published "When AI builds itself" by Marina Favaro and Jack Clark, arguing AI is increasingly designing and shipping its successors: over 80% of code merged at Anthropic in May 2026 was written by Claude and engineers merge about 8x as much code per quarter as in 2024. It sketched three scenarios up to full recursive self-improvement and said Anthropic would join a slowdown if other frontier developers in multiple countries did so verifiably. It was updated in September 2026.

## Key facts
- Code: more than 80% of code merged into Anthropic's codebase in May 2026 authored by Claude; ~8x more code merged per quarter in Q2 2026 than in 2024; Claude shipped 800+ PRs in April 2026 that cut a class of API errors by 1,000x (secondary summaries)
- Task horizon: doubling roughly every four months, from ~4-minute tasks to ~12-hour tasks within a year
- Research: in a bounded, open-ended AI-safety study, Claude-powered agents recovered 97% of the performance gap over ~800 cumulative hours vs 23% for human researchers in one week
- Three scenarios: a capability plateau with wide diffusion; compounding human-supervised efficiency gains; full recursive self-improvement in which models design their successors
- Pause position (Fortune): a slowdown only if 'multiple well-resourced labs at or near the frontier, in multiple countries, agree to stop under the same conditions'; calls for verifiable international coordination
- Quote: 'The doing (i.e., writing the code, running the experiment, producing the result) now costs almost nothing in human time, even if it still has costs in compute.'
- Critics (Fortune) noted the timing, shortly before Anthropic's confidential IPO filing; the September 'R&D Automation Index' turned the argument into measured numbers ([2026-09-17-anthropic-rd-automation-index](https://postcutoff.com/e/2026-09-17-anthropic-rd-automation-index/index.md))

## What happened

Anthropic made its public case that AI-built AI is already under way inside the company, with internal productivity numbers, and stated the
conditions under which it would stop. The essay preceded Dario Amodei's September "Pace the Frontier" campaign and the Institute's R&D
Automation Index.

## Why it matters

It is the clearest statement by a frontier lab that recursive self-improvement is a near-term planning case, together with a concrete,
conditional pause offer. Internal figures are self-reported.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 35 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): in its training data
- Gemini 3.8 Flash (training cutoff March 2026): 65 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 4 days after its cutoff

## Sources

1. [Anthropic Institute: When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement) (anthropic.com, official)
2. [Fortune: Anthropic warns AI could soon improve itself, backs a conditional pause (June 5, 2026)](https://fortune.com/2026/06/05/anthropic-ai-pause-development-recursive-self-improvement) (fortune.com, press)
3. [Forbes (Lance Eliot): Anthropic declares the next big step is AI that builds itself via recursive self-improvement (June 7)](https://www.forbes.com/sites/lanceeliot/2026/06/07/anthropic-recursive-self-improvement-ai/) (forbes.com, press)
4. [MIT Technology Review: AI's recursive self-improvement might not come so quickly after all (Aug 18)](https://www.technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement/) (technologyreview.com, press)

## Changes

- 2026-10-09 (filed): Created (missed at publication, found via the Anthropic Institute page). The secondary-source details (800+ PRs) are from summaries and should be checked against the page.

## Related

- 2026-09-17: [Anthropic's first R&D Automation Index](https://postcutoff.com/e/2026-09-17-anthropic-rd-automation-index/index.md)
- 2026-09-12: [Dario Amodei publishes "We Must Pace the Frontier", calling for a deliberate slowdown](https://postcutoff.com/e/2026-09-12-dario-amodei-pace-the-frontier/index.md)
- 2026-08-28: [Anthropic: automated Claude researchers mitigate 10 alignment failures and nearly match production alignment of an Opus 4.8 checkpoint](https://postcutoff.com/e/2026-08-28-anthropic-automated-alignment-researchers/index.md)
- People: [Dario Amodei](https://postcutoff.com/person/dario-amodei/), [Jack Clark](https://postcutoff.com/person/jack-clark/), [Marina Favaro](https://postcutoff.com/person/marina-favaro/)