METR Report Shows Anthropic Using AI to Accelerate AI Development

The AI safety organization METR published on Thursday (November 5) a report documenting, in more detail than anything seen to date, how Anthropic is using artificial intelligence to build artificial intelligence faster. The investigation lasted 11 weeks and took place inside the company, with access to confidential research data, model performance metrics, and employee interviews.

The central finding: Anthropic's models are improving about twice as fast as they were a year ago. According to METR, benchmark numbers suggest that "Anthropic has made roughly two years of historical progress in the past year." The researchers say they are confident that this acceleration does not simply come from more computing power in training, which has historically been the main driver of progress in AI.

AI Is Already the Biggest Driver of Its Own Evolution

METR states that most of the acceleration recorded through October "is caused by AI," and that this contribution has been growing over time. The conclusion did not come only from benchmark metrics: it also drew on informal Slack polls and interviews with Anthropic employees, who reported significant productivity gains.

Employees using the company's most advanced internal model reported, on average, being about 12 times more productive at work than they would be without AI assistance. That is roughly double what the same employees reported just a few months earlier, using an internal model codenamed Mythos 5.

For now, much of this acceleration still comes from AI assisting human researchers, not acting on its own. But METR warned that "there are some signs that fully autonomous contributions will play a larger role soon," and said it found no clear evidence of diminishing returns that could slow an even greater acceleration.

The Models No One Outside Has Seen

Much of this productivity gain comes from internal models that Anthropic has not yet released publicly. According to METR, the company completed training on three new models in the past six months, in addition to an older model identified in the investigation:

• Model 2 - training completed in July 2026
• Model 3 - training completed between August and September 2026
• Model 4 - training completed between September and October 2026
• Unreleased 2022 - unreleased model, identified by METR with a 2022 date

The names are codenames assigned by METR itself for the investigation; Anthropic uses different internal nomenclature, which was not disclosed. The pace of completing three relevant models in six months gives a concrete sense of how much faster the internal development cycle has become compared to what Anthropic shows to the public.

Half the Report Was Left Out

Anthropic was not obligated to open its data to METR, and, according to the organization itself, it "shared more information than was initially planned for this project, exceeding our expectations." Even so, access fell far short of unrestricted: there was no "as if an employee" access that the company promised to other METR safety researchers under a program announced last month, created after hacking incidents that hit the company in the middle of the year.

The agreement also gave Anthropic the power to redact portions of the final report, and the company used that prerogative extensively. In a post on X, METR itself stated it had delivered a full report of about 100 pages, from which most of the evidence considered most informative was cut from the public version. The organization reinforced that its "information pipeline was relatively narrow," with some data requests denied.

Sam McCandlish, co-founder of Anthropic, publicly acknowledged the extent of the redactions:

We redacted a lot, and I accept that this makes it harder for others to check METR's work. But we determined that sharing exactly what is driving the speedup isn't the right thing to do, both because it's commercially sensitive and because it risks handing out a roadmap for others to go faster.

Sam McCandlish, co-founder of Anthropic

McCandlish said the company will publish its own report on the speed of AI progress "in the coming weeks," along with new safety measures. Anthropic declined to answer The Verge's questions about the report and its fallout, pointing only to that post.

The Criticism: Asking the Industry to Slow Down While Speeding Up

The publication sparked a strong reaction among safety researchers. The point of friction is that Anthropic has built its public image around advocating for responsible development, with CEO Dario Amodei publicly asking the industry to hit the brakes, just as the company prepares to go public. METR's report suggests the opposite in internal practice.

Daniel Kokotajlo, a former OpenAI researcher who now runs the AI Futures Project (responsible for the "AI 2027" forecast scenario), was one of the harshest critics:

recklessly gambling with our lives

Daniel Kokotajlo, director of the AI Futures Project

Kokotajlo also called it contradictory that Anthropic cites competition as justification for the redactions in the report, since the company itself "doesn't seem to be slowing down much."

Other researchers brought the debate into technical territory. Marius Hobbhahn, CEO of Apollo Research (a company that runs safety evaluations for major AI labs), told The Verge that the pace described in the report is even faster than predicted by "AI 2027" itself, a scenario that projected a major advance toward self-improvement in early 2027. Alan Chan, a senior researcher at the GovAI policy center, said the findings should lead skeptics to reconsider the possibility of AI accelerating AI itself, and possibly to shorten their forecasts for when systems will be able to fully automate tasks like software engineering.

Brianna Rosen, director of frontier safety research at the Institute for AI Policy and Strategy, pointed to a governance problem: today it is the companies themselves, not governments, that decide the pace of AI research automation. For her, this undermines the possibility of real democratic oversight over the process.

Why This Matters for Software Builders

The report concerns Anthropic's internal research, but the practical effect lands directly on those who integrate Claude models via API or compete with them day to day. If the speed of evolution has doubled in a year and much of the gain no longer comes from more computing power but from AI optimizing its own research cycle, the interval between versions tends to shrink, and capability jumps tend to become less predictable from public data.

This has a concrete consequence for those building products on top of these models: safety and behavior evaluations done today on one version may lose validity faster than usual, because the next model isn't just "bigger," it's built through a research process that the company itself admits it doesn't fully understand in terms of causes. The opacity around exactly what accelerated progress (the very point of the redactions in the report) makes it harder for outside teams to anticipate behavior changes before a version migration.

There is also a broader backdrop: in September, OpenAI said it was "making strong progress" toward a system capable of advancing deep learning and alignment research more autonomously by March 2028. Combined with METR's report, the target of automation is no longer just writing code: it's AI research itself, including the parts that today ensure a model is safe enough to run in production.

What Remains Unresolved

No regulatory body closely followed this investigation. In the United States, state assemblyman Alex Bores, one of the figures shaping the Democratic agenda on AI, asked Congress to ban recursive self-improvement (RSI) "this year, if possible," a request gaining traction with the arrival of a newly elected Democratic caucus with a more critical stance toward AI companies.

Hobbhahn summed up the mood among safety researchers this way:

Brace for craziness. It is going to go very fast.

Marius Hobbhahn, CEO and founder of Apollo Research

For those who depend on these models to build products, the practical takeaway is less about predicting the distant future and more about work habits: following changelogs and model cards more closely, testing for behavior regressions with every relevant update, and not treating an API version as stable indefinitely just because the version number hasn't changed.