Independent Watchdog Finds Anthropic's AI Models Are Accelerating Their Own Development

Imagine a factory where the machines on the line start helping design the next generation of machines, and that next generation works even faster. That is roughly the dynamic an independent safety group says it is seeing inside Anthropic right now, and it is the kind of finding that should make all of us pay attention, whether we work in tech or not.

METR, an AI safety research organization that specializes in evaluating frontier models, released a risk assessment this week examining how quickly Anthropic's AI systems are advancing their own capabilities. The headline finding is stark in its simplicity. Anthropic is "likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025," according to METR. The group attributes that acceleration to Anthropic's "extensive" use of AI itself for research and development work.

In plain terms, Anthropic appears to be using its own AI tools to build better AI tools, and that feedback loop is speeding things up. Think of it like a construction crew that normally builds one house a month, except now some of the crew members are robots that also help design and assemble the next batch of robots. Each cycle gets a little faster than the last. That is useful when the work is well-defined and the stakes are low. It becomes a much bigger deal when the end product is a technology capable of reshaping labor markets, information ecosystems, and potentially national security.

METR did not produce this report in isolation. The organization says it was in communication with Anthropic starting in mid-August, and it submitted the final public version on November 3rd. That timeline matters, because it places the assessment squarely in the middle of an already tense public conversation about pacing. In mid-September, Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier," in which he argued that AI development needs guardrails to keep up with its own speed. METR's findings now offer some independent, if cautious, evidence for exactly the kind of acceleration Amodei was warning about.

What the Numbers Actually Mean

It is worth pausing on the word "likely" in METR's own phrasing, because responsible risk assessment is rarely about certainty. METR frames its conclusion as a probable trend based on observed patterns, not a definitive, airtight measurement. That caution is itself a sign of good practice. When an outside watchdog is evaluating a company's internal R&D speed, it is working with limited visibility into proprietary systems and internal metrics. The honest answer is often "likely" or "probably," not "certainly."

Still, a doubling in the pace of capability improvement on well-specified tasks is a meaningful signal. Well-specified tasks are the kind of clearly bounded problems researchers use to measure progress, things like coding benchmarks or specific reasoning challenges, as opposed to open-ended, real-world messiness. Progress on these tasks is a reasonable proxy for how fast the underlying technology is maturing, even if it does not capture everything about real-world deployment risk.

Anthropic itself has not stayed silent on this. Co-founder Sam McCandlish said on social media that the company plans to share its own report analyzing the speed of AI progress internally. That is a notable move. It suggests Anthropic recognizes the public interest in this question and wants to offer its own data points alongside METR's independent read, rather than leaving the narrative to outside observers alone. Whether that internal report confirms, complicates, or contradicts METR's twice-the-rate estimate will be worth watching closely.

This all lands inside a broader industry moment that The Verge has been tracking under the banner of an "AI superintelligence slowdown" conversation, a sign that even people building these systems are grappling publicly with how fast is too fast. It is a healthy sign, frankly, that a safety organization can publish findings like this and have the company itself respond with transparency rather than dismissal. That kind of back-and-forth, company and independent evaluator talking past each other through public essays and data releases, is exactly the kind of accountability structure many AI safety advocates have been asking for.

Why It Matters

Here is the honest tension at the heart of this story. Faster AI research and development sounds, on its face, like good news. It could mean faster medical breakthroughs, faster climate modeling, faster solutions to problems that have stumped researchers for decades. But speed cuts both ways. If a company's own AI tools are accelerating that company's ability to build even more capable AI, the normal human checkpoints, the careful testing, the slow deliberation, the chance for outside scrutiny, can get compressed right along with the development timeline.

That is precisely the risk Amodei flagged in his own essay, and it is why an independent assessment like METR's carries weight beyond its specific numbers. Public health and safety fields learned long ago that self-reporting from an interested party is useful but incomplete. You need outside eyes. METR's report, imperfect and probabilistic as it is, functions as one of those outside eyes for an industry that otherwise moves largely behind closed doors.

None of this means Anthropic is acting recklessly, and nothing in METR's findings suggests the company is cutting corners. What it does mean is that the pace of change inside frontier AI labs is now fast enough that even the companies building these systems are commissioning outside checks on themselves, and publicly discussing the results. That is a step in the right direction. The real test will be whether this kind of transparency becomes standard practice across the industry, or remains an exception driven by one company's particular culture of caution.