Anthropic Calls for Options to Pause AI Development

Anthropic has called for options to slow or temporarily pause frontier AI development, arguing that AI systems are increasingly contributing to the development of future AI systems and could eventually become capable of improving their own successors.

The proposal examined the prospect of recursive self-improvement, a scenario in which AI systems automate an increasing share of AI research and engineering work. 

It said recursive self-improvement is not inevitable, but could arrive sooner than many institutions are prepared for. Anthropic presented internal and public benchmark data to support its argument that AI is already accelerating AI development.

The company said more than 80% of the code merged into its production codebase is now authored by Claude, while the typical engineer merged roughly 8 times as much code per day in the second quarter of 2026 as in 2024.

Although Anthropic noted that lines of code are an imperfect productivity metric, it argued that the figures demonstrate a substantial increase in engineering output.

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-2025. pic.twitter.com/QCc9cqGgf4

— Anthropic (@AnthropicAI) June 4, 2026

The report also points to advances in AI-assisted research. Anthropic said Claude has become increasingly capable of executing experiments, debugging complex systems, reviewing code, and solving open-ended engineering problems with limited human intervention. While humans still set research priorities and evaluate results, the company argues that the boundary between human-led and AI-led development is narrowing.

“We also see evidence that people at Anthropic are using Claude to do work that simply wouldn’t have happened otherwise, like building exploratory tooling and addressing long-deferred cleanup,” said Anthropic. “For example, in April 2026, Claude shipped over 800 fixes that reduced a class of API errors by a factor of one thousand. The engineer overseeing Claude estimated that a human would have taken four years to complete this work.” 

According to Anthropic, the key question is whether future systems will develop the judgement required to select important research directions rather than merely execute tasks. 

The company said early internal evidence suggests progress in this area, though significant gaps remain between current models and a system capable of autonomously driving AI research.

Anthropic outlined several potential futures, including one in which AI labs gain compounding productivity benefits from increasingly automated research and development, and another in which AI systems become capable of designing and refining successor models themselves. 

In the latter scenario, humans would primarily oversee, validate and verify the outputs of AI-operated research systems.

Against that backdrop, Anthropic argued that the world should have the option to slow development if advanced AI capabilities begin to outpace governance, safety research or societal adaptation. 

“If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing,” the company wrote.

The paper does not call for an immediate halt to AI development. Instead, Anthropic said it would research verification systems and coordination mechanisms that could make a future slowdown credible and enforceable. 

The company argued that any meaningful pause would require participation from multiple frontier AI developers and governments, along with methods to verify that all parties had actually reduced or halted development.

Anthropic said it plans to convene policymakers, researchers, civil society groups and AI companies to examine questions surrounding recursive self-improvement, coordination and AI governance, and to publish further research in the coming months.

The report comes as Anthropic expands access to its most advanced cybersecurity model, Claude Mythos Preview. This week, the company widened Project Glasswing from roughly 50 initial participants to about 200 organisations across more than 15 countries and has discussed extending access to European institutions through the programme. 

Anthropic says Glasswing partners have already identified more than 10,000 high- and critical-severity software vulnerabilities using Mythos. 

The expansion also follows a string of frontier model releases from the company, including Claude Opus and Sonnet variants that have ranked among the industry's leading coding models, alongside Anthropic's growing focus on autonomous software engineering and cybersecurity workflows.