Anthropic is using AI to help build better AI. That is no longer just a future scenario or a lab slogan. A new assessment from the AI safety organization METR suggests the company’s research is accelerating sharply, while the evidence needed to judge that acceleration remains partly behind closed doors.
That combination is the real story. The question is not only whether AI can speed up AI development. It is whether anyone outside the labs will be able to see, test, and challenge what is happening quickly enough.
AI is moving from assistant to research multiplier
According to the reported METR findings, Anthropic’s models are improving roughly twice as quickly as they were a year ago. The researchers said the company had made approximately two years of historical progress during the last year, and that the acceleration could not be explained simply by adding more computing power.
The reported evidence points instead to AI contributing directly to the research process. Informal Slack polls and employee interviews suggested that Anthropic staff using the company’s most capable internal model considered themselves around 12 times as productive as they would have been without AI assistance. That figure was roughly double the productivity uplift employees had reported a few months earlier while using an earlier internal model known as Mythos 5.
The important detail is that this does not yet describe a fully autonomous AI lab. Much of the reported gain still comes from researchers working with AI systems. But METR reportedly found signs that autonomous contributions could become more important soon, with no clear evidence of diminishing returns or an obvious bottleneck that would stop the acceleration from increasing.
That changes the role of the model. It is no longer only a tool for writing code, summarizing papers, or generating experiments. It becomes part of the mechanism that improves the next research system. The behavior to watch is the feedback loop: better models help researchers move faster, faster research produces better models, and the development cycle becomes less dependent on human capacity alone.
The transparency gap may become the limiting factor
The assessment also reportedly identified three unreleased Anthropic models that completed training between July and October, alongside an older internal model from 2022. Those details offer a rare view into the pace of frontier model development, but they also show how little outsiders can normally see.
Anthropic gave METR access to confidential research, performance data, and employee interviews. It also retained the ability to redact material from the public report. METR reportedly said Anthropic removed a majority of the evidence it considered most informative, while noting that some data requests were denied and that the evaluation pipeline was relatively narrow.
Anthropic’s justification is commercially understandable. Revealing exactly what drives the speedup could give competitors a roadmap for moving faster. But that argument creates a difficult contradiction. A company can say that disclosure is dangerous because it might accelerate the industry, while continuing to accelerate the industry itself.
This is where the story moves beyond Anthropic. Independent evaluation only works when evaluators can explain what they saw, how they reached their conclusions, and where uncertainty remains. If the company being assessed can remove much of the supporting evidence, the public is asked to trust the evaluator’s judgment without being able to inspect the underlying case.
That may be acceptable for a confidential security audit. It is harder to accept when the subject is the speed at which AI systems may begin improving the systems that replace them. The reported METR findings do not prove that recursive self-improvement is imminent, and they do not establish that Anthropic’s internal productivity figures are representative of the wider industry. They do show why the old distinction between AI product development and AI research is becoming unstable.
For companies, policymakers, and researchers, the strategic consequence is straightforward: progress may be speeding up faster than the systems designed to oversee it. The next contest will not be only about who builds the most capable model. It will also be about who can produce credible evidence that the model’s development remains understandable, challengeable, and governable.