Anthropic’s AI progress is speeding up — and that could be the problem

Anthropic is developing its AI systems roughly twice as fast as it was a year ago, according to a new METR report published this week. The finding matters because it suggests the safety-minded lab is accelerating the very frontier technology its chief executive has publicly urged the industry to slow down.

What METR found inside Anthropic

The report, described by The Verge and also summarized by Deccan Chronicle, says Anthropic’s model improvement has roughly doubled compared with a year earlier. METR, an AI safety research group that studies advanced systems and lab practices, said the pace of progress was not just a matter of throwing more compute at the problem. Instead, the group concluded that AI itself now appears to be doing more of the work behind the speedup.

In METR’s framing, Anthropic made about two years of historical progress in the last year. That is a striking claim on its own, but the report goes further: it says employees using Anthropic’s most capable internal model self-reported around 12 times the productivity compared with their baseline work. The report also says it found no clear sign that diminishing returns or bottlenecks were slowing the acceleration.

The public version of the report was heavily narrowed by redactions. According to The Verge, METR said Anthropic redacted a majority of the most informative evidence, and that some data requests were denied. Access to the company’s internal information was described as relatively narrow, which limits how independently outsiders can check the underlying numbers and methods.

How did Anthropic get here?

METR said Anthropic completed training three unreleased models in the past six months, with the work moving quickly across the summer and early fall. The report places those milestones as follows:

The report itself was based on an 11-week investigation inside Anthropic. By October 2026, METR said the majority of the acceleration was being driven by AI rather than merely by larger datasets or more hardware. That distinction matters for the broader industry because it points toward a feedback loop: AI helping build better AI, a concept often described as recursive self-improvement.

In other words, this is not just a story about a faster research lab. It is a story about how the tools are changing the pace of the lab itself.

Why it matters

This finding lands at the center of one of the most uncomfortable debates in artificial intelligence: whether frontier labs can reliably judge the risks of the systems they are racing to improve. Anthropic is known for emphasizing safety, responsible deployment, and public discussion of AI dangers. Yet the report suggests the company is also speeding up its own development substantially, even as its leadership argues publicly for restraint.

That tension is why the report has drawn so much attention. If a major safety-focused lab is using AI to accelerate progress, then the question is no longer whether AI will be used to speed up R&D. The question is who gets to see how that acceleration works, how fast it is moving, and whether outside oversight can keep up.

The policy stakes are equally clear. Brianna Rosen, research director for frontier security at the Institute for AI Policy and Strategy, argued that the real power may already sit with companies rather than governments when it comes to deciding how research is automated. Her point echoes a broader concern in AI governance: if the most advanced labs control both the systems and the evidence about their effects, public oversight becomes much harder.

There is also an economic angle. If internal AI tools are multiplying productivity as METR reports, then frontier labs may be able to compress years of work into months. That could intensify competition among major AI companies, widen the gap between leaders and followers, and make transparency even more elusive as firms guard methods they see as strategically valuable.

What exactly did the report suggest about speed?

The report suggests the acceleration is real, measurable, and faster than many outside observers expected. METR said Anthropic’s model improvement is now happening about twice as fast as it was a year ago. It also said the company appears to have made roughly two years of historical progress in a single year, which is a blunt measure of how quickly capabilities are compounding.

Those findings were echoed in broad terms by Deccan Chronicle, which also reported that Anthropic is developing AI about twice as quickly as it was a year ago. The overlap between the outlets strengthens the headline claim, even if the details are more fully developed in The Verge’s account of the METR report.

But speed alone is not the whole story. The report’s more important implication is that AI may now be accelerating the research process itself. If that is true, the industry may be entering a phase where the bottleneck is less about raw compute and more about how effectively AI systems can be used to automate scientific and engineering work.

By the numbers

• Roughly 2x faster model improvement than a year earlier
• About 2 years of historical progress made in the last year
• Around 12x employee self-reported productivity using Anthropic’s top internal model
• 3 unreleased models trained in the past six months
• 11 weeks for METR’s investigation inside Anthropic

Why are researchers worried about the redactions?

The concern is that the public cannot fully inspect the evidence behind the headline. According to The Verge, METR said Anthropic redacted a majority of the most informative evidence from the report and denied some data requests. That makes the headline result harder to independently verify, even if the core conclusion remains clear.

Sam McCandlish, Anthropic’s cofounder, defended the limited disclosure by saying the company believed some details were commercially sensitive and could aid rivals. In one of the report’s most revealing comments, he said,

“We redacted a lot, and I accept that this makes it harder for others to check METR’s work”

He also said,

“sharing exactly what is driving the speedup isn’t the right thing to do”

That position captures the central conflict in the story. The lab wants to avoid giving competitors a blueprint for its own acceleration. Critics argue that the same secrecy also weakens democratic oversight and leaves the public to trust the lab’s own judgment on safety at a moment of rapid change.

How are critics reading the findings?

Many critics see the report as evidence that Anthropic is speeding up the same technology it warns could become dangerous. That criticism is not subtle: the report sharpens the broader concern that safety rhetoric can coexist with rapid capability gains.

Brianna Rosen said the findings reinforce the need for independent evaluators to have enough freedom to explain their conclusions. Her argument is straightforward: if outside experts cannot explain why they reached a conclusion, then public scrutiny becomes little more than a stamp on a process controlled elsewhere.

Marius Hobbhahn, CEO and founder of Apollo Research, said the pace appears even faster than the widely discussed AI 2027 forecast. In his words,

“Brace for craziness”

Alan Chan said the findings should cause skeptics to rethink timelines for fully automated software engineering. That is a serious claim because software engineering is one of the clearest near-term test cases for AI automation. If AI systems are already helping labs speed up their own development work, the leap from assistance to autonomy may be smaller than some expected.

Daniel Kokotajlo went further, accusing Anthropic of recklessly gambling with lives. That is a severe charge, and it reflects just how high the stakes have become in the debate over frontier AI development. For many critics, the issue is not simply whether Anthropic is moving fast. It is whether any institution can move this fast and still claim it is adequately managing the risks.

Background: why Anthropic’s position is so sensitive

Anthropic occupies a highly visible place in the AI industry. It is a major frontier AI lab, and its public identity is built around safety and responsible deployment. Its chief executive, Dario Amodei, has publicly urged slowing AI development, which makes the METR findings especially awkward for the company.

That tension is part of a larger pattern across frontier AI. The most advanced companies often say they are carefully managing risk while continuing to push model capability forward as quickly as possible. Their systems are too strategically valuable to reveal in full, yet their risk profiles are too important to ignore. The result is a constant push and pull between secrecy and scrutiny.

METR’s role also matters. The group is not a consumer watchdog or a market analyst. It is an AI safety research organization that evaluates advanced systems and lab practices. That means its report carries a different kind of weight: it is not merely describing performance, but probing the machinery of progress itself.

In that sense, the report is less a single-company story than a warning about the structure of the field. If AI systems are increasingly helping build better AI systems, then the pace of progress may be driven by a loop that becomes harder to observe as it accelerates. The more automated the process becomes, the less clear it may be who can meaningfully audit it.

What happens next

Anthropic says another piece of the story is still coming. McCandlish said the company will publish its own report on AI progress in the coming weeks, which could give the public a fuller account of what is driving the speedup. That report may also show whether Anthropic is willing to disclose any additional safety steps, or whether competitive concerns will keep the details limited.

For now, several open questions remain:

• How much of METR’s evidence was redacted, and which claims are hardest to verify?
• What will Anthropic’s own report say about the source of the acceleration?
• Will the company disclose any new safety measures beyond McCandlish’s explanation?
• Will independent evaluators gain broader access to frontier labs in the future?
• How quickly will AI-assisted research move from human-augmented work to fully autonomous contributions?

The next few weeks should show whether Anthropic’s planned follow-up report answers any of those questions or simply restates the company’s preference for limited disclosure. Either way, the METR findings have already reframed the debate. The issue is no longer whether AI can accelerate AI research. It can. The real question is whether the industry can explain that acceleration clearly enough for everyone else to judge the risks.

And if the answer remains no, the gap between technical progress and public accountability may widen faster than the technology itself.

Timeline

Note: The report’s public conclusions are clear, but its most detailed evidence was heavily redacted, making independent verification more difficult than usual.