AI Is Helping Anthropic Build Better AI Faster. The Details Are Secret.
AI Is Helping Anthropic Build Better AI Faster. The Details Are Secret.
There has always been a slightly unnerving idea buried inside the artificial intelligence boom: What happens when the technology being developed becomes useful enough to help develop the next generation of itself?
We may be beginning to find out.
In an unusually revealing assessment published November 5, METR, an independent organization that studies advanced AI systems, reported that Anthropic appears to be improving the capabilities of its models at roughly twice its previous rate. From August 12 through October 29, METR researchers conducted what they call a high-access evaluation of Anthropic's internal AI research-and-development operation, examining model benchmarks, training information, AI-speedup analyses, forecasts and interviews with employees. The access was broad but not unrestricted: METR says it did not have employee-like, unmediated access to Anthropic's internal data, documents or models, did not sample non-public models, and could have information requests rejected by Anthropic. Their job was to determine whether AI was beginning to accelerate AI research itself.
Their answer, with considerable uncertainty attached, was yes.
Measured using Anthropic's version of the Epoch Capabilities Index, the company's internal model benchmarks have been improving at an estimated 1.9 times their earlier rate since October 2025. METR describes that another way: Anthropic appears to have made roughly two years' worth of its previous benchmark progress during the past year. The researchers put the 95 percent confidence interval around that estimate at 1.5 to 2.8 times the earlier rate. That doesn't mean Anthropic's AI has become useful in the real world twice as fast. AECI measures progress on well-specified tasks, and METR cautions that the result may not translate cleanly to broader, messier work.
The obvious explanation might be that Anthropic simply threw dramatically more computing power at the problem. METR says that's not what its evidence showed. More precisely, the researchers say they received convincing evidence that the acceleration wasn't caused by an acceleration in the scaling of training compute. Instead, their assessment is that a majority of Anthropic's current acceleration is probably attributable to its use of AI in its own R&D, although METR stresses that there is substantial uncertainty around that judgment.
That's where an otherwise technical report becomes a much bigger story.
When the Tool Becomes Part of the Factory
The concept is simpler than the terminology surrounding it. Anthropic builds increasingly capable AI systems. Its researchers then use those AI systems to help them conduct research, write code, analyze results and perform other work involved in developing AI. If those researchers become more productive, that could help the company produce the next generation of models faster. Those improved models could then assist with another round of development.
The possible result is a feedback loop: better AI helps humans develop better AI, which helps humans develop still better AI.
METR isn't claiming that Anthropic has reached some science-fiction moment in which autonomous machines have taken control of their own development. In fact, the report specifically says researchers were told that today's acceleration comes primarily from AI assistance, rather than AI systems making fully autonomous contributions to the most important parts of the training process. But researchers also reported seeing what they called “signs of life” suggesting that more autonomous contributions could emerge, with considerable variation depending on the area of research.
The human researchers themselves appear to be feeling the difference. METR interviewed Anthropic employees about how much AI was accelerating their work and found a wide range of estimates, but said the responses centered around approximately a twofold speedup. Anthropic also supplied internal self-reported productivity information, including informal Slack polls. METR appropriately warns against taking those numbers too literally, calling the polls only “very loose evidence.” Those individual productivity estimates shouldn't be confused with Anthropic's overall rate of AI progress: METR explicitly warns that a researcher's reported speedup does not translate directly into an equivalent acceleration in company-wide capabilities progress. Still, taken alongside the benchmark data and other evidence, the interviews helped support the broader picture the investigators were seeing.
And then we reach the uncomfortable part of the report.
The Evidence We Aren't Allowed to See
METR got unusually broad access to Anthropic. The public did not get unusually broad access to what METR found.
Under the terms of the assessment, Anthropic retained the unilateral right to redact non-public information before publication. METR, however, retained the right to tell readers that redactions had occurred and, critically, whether its findings depended significantly upon information that had been removed.
Anthropic used those redaction rights extensively.
METR says Anthropic removed “a majority of the evidence we found most informative” for reaching its conclusions. The researchers were still allowed to publish their highest-level judgments, but substantial evidence underlying those judgments remains invisible to anyone outside the evaluation.
The researchers saw the evidence. Anthropic saw the evidence. We get the conclusion.
That creates an extraordinary situation. An independent organization has effectively been permitted to tell the public what it concluded after examining the inner workings of one of the world's frontier AI developers, while being prohibited from showing the public much of what convinced it.
To its credit, METR doesn't pretend this is satisfactory. The organization explicitly says the redactions make it harder for outsiders to evaluate the evidence, compare it with information from elsewhere and reach their own conclusions. As a result, METR says the public must rely on its judgment “much more than would be ideal.”
That raises the question hanging over the entire assessment: What, exactly, did Anthropic not want published?
METR can't tell us the contents, because that's precisely the material that was redacted. But the public report does identify the categories.
The withheld material includes key intermediate measurements of internal capabilities, estimates of algorithmic progress, measurements of important R&D inputs, portions of evidence concerning AI's contribution to research, analyses of specific bottlenecks and forecasts. METR also says it couldn't publish its more detailed estimates of acceleration across important portions of Anthropic's model-training pipeline, a fuller breakdown of the roles played by compute, data and algorithmic progress, or forecasts of metrics for future models.
Even more strikingly, METR says some of the redacted evidence caused its researchers to take the observed acceleration more seriously. That evidence offered some indication that the rate might be greater than 2x and increasing over time, though METR emphasizes substantial uncertainty and says it would be surprised if the acceleration were significantly higher.
That doesn't tell us what's behind the black bars. But it tells us why those black bars matter.
The Case for Secrecy
There is a perfectly serious reason Anthropic might not want this information made public. According to METR, Anthropic said its redactions were “aimed at removing confidential information that could accelerate other AI developers.”
That explanation shouldn't be waved away. If Anthropic has information about its research process that could help other AI developers accelerate their own work, publishing it could expose competitively sensitive information. A report intended to measure whether AI is accelerating AI research could, perversely, become a document that helps accelerate AI research elsewhere.
Nor does METR claim that every piece of hidden evidence points toward some dramatic breakthrough. Quite the opposite. The researchers say much of the redacted material has multiple plausible interpretations and acknowledge disagreements and uncertainty about how the evidence should be assembled into a coherent picture.
So “Anthropic redacted the report” cannot honestly be translated into “Anthropic is hiding a dangerous discovery.” We don't have evidence for that conclusion.
Something more interesting is happening.
Anthropic's explanation creates a genuine tension between competitive secrecy and public accountability. The more valuable AI becomes as a tool for developing AI, the more commercially sensitive information about that process becomes. Yet the more rapidly AI begins accelerating its own development, the more important that same information becomes to researchers, governments and the public trying to understand where the technology is going.
The more consequential AI-assisted AI development becomes, the stronger the incentives may become to keep its most revealing details secret.
Those interests point in opposite directions.
The Real Experiment Has Already Started
For years, recursive AI development has largely been discussed as a future possibility: the moment when increasingly capable machines begin contributing meaningfully to the production of their successors.
METR's report suggests we should be careful about imagining that transition as a single dramatic event.
The mundane version may arrive first.
A researcher delegates some coding to an AI assistant. Another uses it to analyze an experiment. Another becomes more productive. Models improve. The next model becomes a better research assistant. If improvements like these accumulate across an organization, they could create the kind of accelerating development cycle METR is trying to measure.
That is broadly the phenomenon METR went looking for at Anthropic. The investigators believe they found meaningful evidence of it. They estimate that Anthropic's capability progress on well-specified tasks is running at around twice its earlier rate, believe AI probably accounts for a majority of the current acceleration, and say they didn't identify convincing bottlenecks that would clearly prevent further rapid progress during the next 12 months, although they emphasize considerable uncertainty around that last assessment.
And that's ultimately why the redactions matter.
The provocative question is “What is Anthropic hiding?”
The more important question may be “What happens when the information society most needs to evaluate AI progress becomes the same information AI companies have the strongest incentive to keep secret?”
METR's assessment doesn't answer that question. In a peculiar way, the report itself demonstrates it.
The researchers saw the evidence. Anthropic saw the evidence. We get the conclusion.
And for now, that's where the window closes.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, and AI policy.