AI Is Accelerating Anthropic's Research Behind Closed Doors

In Brief

AI safety nonprofit METR says Anthropic made roughly two years of progress in one year, largely because AI helped its researchers.

Much of the supporting evidence is redacted, leaving businesses with a consequential finding about the pace of change and limited ability to examine how it happened.

What Happened

Anthropic, the company behind Claude, is using AI to speed up its research, according to a new assessment from METR. Evaluators examined research data and interviewed Anthropic staff during a high-access evaluation from Aug. 12 to Oct. 29, 2026. They concluded that the company had compressed roughly two years of progress into the past year.

The finding comes with an important distinction. The acceleration primarily involved AI assisting human researchers. The public report does not establish that AI is independently creating successive generations of more powerful models, a process known as recursive self-improvement that could further accelerate development.

It also comes with a substantial disclosure problem. METR submitted a 100-page report to Anthropic, but the company redacted much of the evidence in the public version. In its explanation of the redactions, METR said readers would have to rely on its judgment more than it considered ideal.

SmarterX founder and CEO Paul Roetzer broke down why faster AI research matters for companies still learning to use the technology, on Episode 250 of The Artificial Intelligence Show.

The Key Numbers

~2 - Years of Anthropic's estimated research progress in the timespan of one year, says METR

Aug. 12-Oct. 29, 2026 - METR's evaluation period inside Anthropic

100 - Number of pages in METR's full report submitted to Anthropic

Faster AI Research Alters Business Planning

The time available to adapt could shrink. Roetzer's immediate answer to why the report matters was blunt: "Your company could get disrupted way faster," he says. In his interpretation, the significance extends beyond a research lab's productivity. Faster development could bring new capabilities into businesses before their people and processes are ready to absorb them.

That puts pressure on assumptions about how much time organizations have to prepare. A company may still be exploring today's tools while the labs are already using AI to help develop what comes next. Roetzer sees a growing need to understand the technology across careers, organizations, and industries.

Human involvement remains central to the finding. METR's assessment supports a narrower conclusion than a quick runaway cycle of AI improving itself. The researchers still matter now.

The evidence is difficult to inspect. Anthropic co-founder Sam McCandlish said many of METR's conclusions matched his own. He defended withholding details because they were commercially sensitive and could provide other developers with a roadmap to accelerate their work. That explanation leaves the public depending heavily on the evaluator's judgment.

"If you think it's moving fast and hard to keep up with now, it's only going to get harder."

— Paul Roetzer, founder and CEO of SmarterX, Episode 250 of The Artificial Intelligence Show

SmarterX Take

The report pushes organizations to examine how prepared they are for capabilities that could arrive sooner than expected. Roetzer's concern applies to the people learning AI as well as the leaders deciding how it will affect their businesses. "We just need to be honest with ourselves about where we are right now," he says. 

That honesty also applies to the evidence. METR's finding is a reason to take acceleration seriously. It cannot tell a business exactly when its work will change, and it does not prove autonomous self-improvement. But leaders can still prepare for faster change.

What to Watch

Anthropic has promised more of its own analysis. McCandlish said the company planned to share a report in the coming weeks on the speed of its AI progress, its position under its Responsible Scaling Policy, and additional safety and security measures. Whether it provides more inspectable evidence will matter.

Human assistance and autonomous research need separate scrutiny. Future assessments should help clarify how much of the work AI can do independently and where researchers remain essential. This should shape how confidently anyone can project the next stage of development.

Can Your Team Keep Up With Faster AI Development?

A lack of education and training is an AI adoption barrier for 38% of respondents to the 2026 State of AI for Business Report. That finding gives Roetzer's warning a practical context: preparing for faster AI development requires people who have the time and support to understand and apply the technology.

Based on more than 2,100 responses from professionals across roles, functions, and industries, the report examines organizational readiness, adoption, governance, and training. Use it to compare your organization's preparation with what other professionals report experiencing and identify where learning needs more attention. Read the full report →