The key question is no longer whether AI will improve itself, but when—and whether institutions are ready.

Claude writes most of its own code

The most striking internal metric is the degree to which AI has already taken over the physical act of coding at Anthropic. As of May 2026, more than 80% of all code merged into Anthropic's codebase is written by Claude . This is a jump from low single-digit levels a year prior, in February 2025 . Earlier, Anthropic CEO Dario Amodei had noted that over 90% of code for new Claude models and features was AI-authored, and the new report confirms that this dominance extends across the entire company's merged codebase .

This is not just an efficiency gain; it represents a structural shift in the development pipeline. The consequence is that Anthropic engineers are now shipping eight times more code per quarter compared to the 2021–2025 baseline . The bottleneck in software creation is shifting from writing and testing to higher-level goal-setting, architecture, and judgment.

Benchmark performance is saturating at an unprecedented pace

The report provides a timeline of breakneck acceleration across several standard measures of engineering and research competence :

• SWE-bench (real-world software engineering): In late 2023, models scored around 2%. By the time of the report, Claude was scoring 93.9%, effectively saturating a benchmark designed to test human-level bug-fixing on real GitHub issues .
• CORE-Bench (reproducing published research): AI success rates jumped from around 20% in 2024 to saturating the benchmark just fifteen months later .
• AI training code optimization: This internal benchmark, which asks a model to optimize code that trains a small language model, saw a staggering leap. In May 2025, the mean speedup was 2.9 times. By April 2026, with the unreleased Claude Mythos Preview, it was 52 times . A skilled human working alone would need four to eight hours to achieve a 4x improvement; the current generation of Claude models exceeds 50x .

The clock speed of autonomy is doubling every four months

Perhaps the most impactful metric for predicting recursive self-improvement is the duration of autonomous tasks. Research from METR tracks how long an AI can work independently at a 50% success rate. This task horizon has expanded from roughly 30 seconds in 2022 to 12 hours with Claude Opus 4.6 by April 2026, a 1,440-fold increase . Claude Mythos Preview can work for at least 16 hours, which is noted as being near the upper limit of what METR can currently measure . The doubling rate for this horizon has accelerated from every seven months to every four months .

A third of AI-assisted work was previously impractical

The quantitative data on code and benchmarks is coupled with internal polling on human productivity. Analysis of 200,000 internal Claude transcripts and 53 in-depth interviews found that 27% of AI-assisted tasks were jobs employees simply would not have attempted without AI, because the time cost previously made them impractical . This is not automation of existing work but an expansion of what is possible to attempt at all. In a separate November 2025 internal study, employees reported using Claude in 60% of their work and estimated a 50% productivity boost, up from 20% the prior year .

Slowing down or containing the risk: a growing rift

Anthropic's position is explicit. The company states, "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for" . It argues that a global ability to pause or slow AI development would "likely be a good thing" and is directly urging other labs to consider it .

The same week, OpenAI's actions painted a contrasting picture. On June 3, OpenAI published a public policy agenda calling for a federal frontier AI safety framework with mandatory model evaluations and whistleblower protections, but also with a critical clause: the preemption of state-level safety laws . It explicitly asks the federal AI safety institute CAISI to prioritize monitoring progress toward recursive self-improvement . Simultaneously, OpenAI is staffing up for this exact risk, creating a "Researcher, Recursive Self-Improvement Preparedness" role within its safety team with a posted compensation of $295,000 to $445,000 . The job is framed as a loss-of-control containment problem, a "tasteful and strategic" effort to mitigate risks that "might exist in the future, but might not exist now" .

Both labs see the same wave approaching, but Anthropic is urging the fleet to slow down, while OpenAI is hiring lifeguards and arguing against any single state issuing a swimming ban.

The three steps toward a self-building AI

The blog post outlines a three-phase trajectory toward the ultimate risk :

1. Human-directed acceleration (now): AI assists human engineers, writing most code, but humans set goals and exercise judgment.
2. Autonomous research and engineering: AI can execute well-specified experiments end-to-end without human guidance on methods; a gap remains on high-level goal-setting and judgment.
3. Full recursive self-improvement ("closing the loop"): AI autonomously designs, builds, and trains its own successor with minimal or no human oversight.

Anthropic co-founder Jack Clark has separately estimated a 60% probability that this "loop" closes by the end of 2028 . The internal data in the June 4th post provides the factual underpinning for why that estimate is not a distant hypothetical but a projection from a curve that is already visibly bending upward.

The key question is no longer whether AI will improve itself, but when—and whether institutions are ready.

Claude writes most of its own code

The most striking internal metric is the degree to which AI has already taken over the physical act of coding at Anthropic. As of May 2026, more than 80% of all code merged into Anthropic's codebase is written by Claude . This is a jump from low single-digit levels a year prior, in February 2025 . Earlier, Anthropic CEO Dario Amodei had noted that over 90% of code for new Claude models and features was AI-authored, and the new report confirms that this dominance extends across the entire company's merged codebase .

This is not just an efficiency gain; it represents a structural shift in the development pipeline. The consequence is that Anthropic engineers are now shipping eight times more code per quarter compared to the 2021–2025 baseline . The bottleneck in software creation is shifting from writing and testing to higher-level goal-setting, architecture, and judgment.

Benchmark performance is saturating at an unprecedented pace

The report provides a timeline of breakneck acceleration across several standard measures of engineering and research competence :

• SWE-bench (real-world software engineering): In late 2023, models scored around 2%. By the time of the report, Claude was scoring 93.9%, effectively saturating a benchmark designed to test human-level bug-fixing on real GitHub issues .
• CORE-Bench (reproducing published research): AI success rates jumped from around 20% in 2024 to saturating the benchmark just fifteen months later .
• AI training code optimization: This internal benchmark, which asks a model to optimize code that trains a small language model, saw a staggering leap. In May 2025, the mean speedup was 2.9 times. By April 2026, with the unreleased Claude Mythos Preview, it was 52 times . A skilled human working alone would need four to eight hours to achieve a 4x improvement; the current generation of Claude models exceeds 50x .

The clock speed of autonomy is doubling every four months

Perhaps the most impactful metric for predicting recursive self-improvement is the duration of autonomous tasks. Research from METR tracks how long an AI can work independently at a 50% success rate. This task horizon has expanded from roughly 30 seconds in 2022 to 12 hours with Claude Opus 4.6 by April 2026, a 1,440-fold increase . Claude Mythos Preview can work for at least 16 hours, which is noted as being near the upper limit of what METR can currently measure . The doubling rate for this horizon has accelerated from every seven months to every four months .

A third of AI-assisted work was previously impractical

The quantitative data on code and benchmarks is coupled with internal polling on human productivity. Analysis of 200,000 internal Claude transcripts and 53 in-depth interviews found that 27% of AI-assisted tasks were jobs employees simply would not have attempted without AI, because the time cost previously made them impractical . This is not automation of existing work but an expansion of what is possible to attempt at all. In a separate November 2025 internal study, employees reported using Claude in 60% of their work and estimated a 50% productivity boost, up from 20% the prior year .

Slowing down or containing the risk: a growing rift

Anthropic's position is explicit. The company states, "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for" . It argues that a global ability to pause or slow AI development would "likely be a good thing" and is directly urging other labs to consider it .

The same week, OpenAI's actions painted a contrasting picture. On June 3, OpenAI published a public policy agenda calling for a federal frontier AI safety framework with mandatory model evaluations and whistleblower protections, but also with a critical clause: the preemption of state-level safety laws . It explicitly asks the federal AI safety institute CAISI to prioritize monitoring progress toward recursive self-improvement . Simultaneously, OpenAI is staffing up for this exact risk, creating a "Researcher, Recursive Self-Improvement Preparedness" role within its safety team with a posted compensation of $295,000 to $445,000 . The job is framed as a loss-of-control containment problem, a "tasteful and strategic" effort to mitigate risks that "might exist in the future, but might not exist now" .

Both labs see the same wave approaching, but Anthropic is urging the fleet to slow down, while OpenAI is hiring lifeguards and arguing against any single state issuing a swimming ban.

The three steps toward a self-building AI

The blog post outlines a three-phase trajectory toward the ultimate risk :

1. Human-directed acceleration (now): AI assists human engineers, writing most code, but humans set goals and exercise judgment.
2. Autonomous research and engineering: AI can execute well-specified experiments end-to-end without human guidance on methods; a gap remains on high-level goal-setting and judgment.
3. Full recursive self-improvement ("closing the loop"): AI autonomously designs, builds, and trains its own successor with minimal or no human oversight.

Anthropic co-founder Jack Clark has separately estimated a 60% probability that this "loop" closes by the end of 2028 . The internal data in the June 4th post provides the factual underpinning for why that estimate is not a distant hypothetical but a projection from a curve that is already visibly bending upward.