Why would an AI company headed for an IPO with a $1T valuation want to hit the brakes?

As a company widely expected to pursue an IPO at a valuation approaching $1 trillion, it might seem odd that Anthropic is contemplating whether development within its industry should be stopped.

But the Claude creator has published research questioning whether the technology it is building should be slowed down or paused, at least temporarily, for the good of society.

The detailed analysis lays out how quickly AI itself has become capable of developing AI, how much further that could go, and why the destination carries risks serious enough to warrant global coordination.

The central issue is what researchers call recursive self-improvement, or the point at which an AI system becomes capable of designing and training its own successor without any meaningful human involvement. Anthropic says that threshold has not yet been crossed but warns it could arrive well before governments, regulators, or businesses are prepared to respond.

"We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for," states the paper, co-authored by Marina Favaro and Jack Clark with a sizeable editorial team.

Is AI building itself?

The range of tasks that AI models can complete independently has been doubling roughly every four months, the research shows; faster than the already rapid trend recorded the year before.

In early 2024, Claude could handle tasks requiring around four minutes of human effort, but by early 2025, that had stretched to around 90 minutes. A year later, it had reached 12 hours. If the pattern holds, tasks that would occupy a skilled professional for days could fall within AI's reach before the end of this year.

Inside Anthropic itself, the average Anthropic employee was handling eight times as much output per day in the second quarter of 2026 compared to 2024 because AI was doing the bulk of the execution while humans focused on direction and oversight.

The paper says that Anthropic’s AI went from matching a competent junior professional's output roughly a year ago to substantially outperforming skilled human researchers today, at a fraction of the time and cost.

Pause can’t be unilateral

"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 paper states. "But if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe."

The paper says the firm would be prepared to slow or pause frontier AI development if other leading companies and governments agreed to the same conditions in a way that could be independently verified. It also concedes the difficulty of achieving that.

Unlike physical weapons programmes, AI development leaves a far lighter footprint, the materials involved are widely available, and the competitive incentive to continue while others stop is enormous.

The paper sets out three plausible paths from here:

• the current pace slows on its own before AI can direct its own development;
• AI handles most execution work while humans retain strategic judgment, multiplying what each person or organisation can accomplish;
• AI begins designing its own successors without human input, at which point progress is limited only by energy and computing capacity.

Anthropic is candid about which future it considers most likely in the near term. Every capability it has tracked, including judgment-intensive tasks, not just routine processing, has continued to improve on the same trajectory without any sign of levelling off.

The company regards the second scenario as the probable destination for the coming years and treats the third as a genuine rather than remote possibility.

What the paper does not offer is any suggestion that the institutions governing financial markets, corporate activity, or professional services are ready for the speed of change being described.

"The window to investigate the questions together is here, and people outside AI companies should be involved in this deliberation," it concludes.