Anthropic says slow AI sprints as it files a confidential IPO

Anthropic just made a very specific ask of the AI industry: slow down, or even temporarily pause, frontier model development. The company said it would be “good for the world” to have that option, according to a blog post written by Anthropic co-founder Jack Clark and researcher Marina Favaro. And yes, the timing is doing a lot of work here. The same week Anthropic began the process of going public with a confidential IPO filing, it also argued that the industry may need to hit the brakes long enough for “societal structures and alignment research” to catch up with the technology.

That is not a small idea, and it is not framed as a casual suggestion. The post says any actual pause would require something closer to nuclear-accord level coordination: agreement from all frontier AI labs plus support from policymakers around the world. Even then, Anthropic says there is no guarantee everyone would play along. The company wrote that “training runs are far easier to conceal than missile silos, their inputs are general-purpose, and the incentive to defect quietly is enormous, because whoever continues while others pause could inherit the lead. A credible pause also has to specify what triggers it, what lifts it, and who adjudicates.” In plain English, this is a coordination problem with cheating baked into the design.

That tension is what makes Anthropic's message so notable. The company has long positioned itself as the more safety-minded alternative to OpenAI, where its cofounders came from. It has also been one of the more alarmist voices in the industry about the risks of increasingly capable AI. Skeptics can, fairly, read some of that as positioning. After all, one easy way to sell expensive enterprise AI is to argue that the same tools could eventually become dangerous enough to warrant caution. But Anthropic is not making this point from the sidelines. It is doing it while preparing for public-market scrutiny and after announcing last week that it reached a $965 billion valuation, putting it among the most valuable companies in the world.

That makes the messaging feel, as the source notes, pretty rich. A nearly trillion-dollar company telling the rest of the field to slow down while it heads toward an IPO is the kind of contradiction that gets boardrooms talking. Still, the surrounding environment suggests Anthropic is not speaking into a vacuum. US President Donald Trump signed an executive order this week that in part directs the Treasury Department to establish an “AI cybersecurity clearinghouse” to work with the AI industry and critical infrastructure operators. The goal is to coordinate and deconflict the use of advanced AI tools. Meanwhile, public backlash to rapid datacenter expansion is growing. So even if no one is about to announce a global pause, the policy weather is getting colder around unrestrained acceleration.

Inside the paper, Anthropic tries to ground the warning in technical trends rather than philosophy alone. The authors say the “human role is narrowing” in model development, and they argue that models could eventually self-improve and write better versions of themselves without people in the loop. Their example is blunt: “Once human- and AI-authored code quality reach parity, humans will stop writing code entirely, and shift to only reviewing it.” But that creates a new constraint. “If they can’t review code as quickly as Claude can generate it, human review will become the bottleneck to AI development.” In other words, the limiting factor stops being the machine's output and becomes the speed at which humans can check it.

Anthropic says the trend is already visible in its own workflow. As of May, Claude authored more than 80 percent of the code merged into Anthropic's codebase, up from the low single digits before Claude Code launched in research preview in February 2025. The company also says newer models are improving on complex tasks faster than before. The length of human tasks models can reliably complete on their own had been doubling every seven months as measured in March 2025. Now, Anthropic says, that pace is closer to every four months. The company also offered a striking comparison: Claude 3 Opus, released in March 2024, could reliably complete tasks that take humans four minutes. Claude Opus 4.6 can reliably complete tasks that take humans 12 hours, the team wrote. And if the trend holds, the paper says, tasks that take a skilled person days could come into range this year.

Anthropic does not claim that all of this is guaranteed. The paper explicitly says unknown bottlenecks could emerge and slow progress, and it points to Amdahl’s Law, the idea that speeding up one part of a system eventually exposes bottlenecks somewhere else. Anthropic says it has already seen one version of that effect: “as we’ve begun to push more code around the organization, human code review has become a new bottleneck.” The authors also say AI still struggles with “taste,” meaning choosing the next step when a human does not prompt it. “Without that judgment, Claude is a capable assistant, but not a system that could drive AI progress on its own,” the paper states. “It is genuinely unclear whether today’s training methods and architectures could unlock that capacity.”

But the real fork in the road is the possibility that the current trend simply keeps holding. In that scenario, Anthropic says AI progress becomes limited mainly by compute, while humans move into oversight, validation, and verification inside an expanding “virtual lab” run by AI systems. That would be a massive shift in how frontier AI gets built, audited, and controlled. For executives, investors, and policymakers, the takeaway is simple: the pace question is no longer academic. If Anthropic is right, the bottleneck is moving from model quality to human governance, and the companies that can prove control, not just capability, may be the ones that can keep scaling without getting their own brakes ripped out by regulators, customers, or their rivals.