Anthropic is reportedly improving its models at roughly twice the rate of a year ago, and the majority of this acceleration does not appear to stem from an increase in compute power used for training. That is according to a new evaluation by METR, an artificial intelligence safety research organization, conducted over an eleven-week period inside the US company.
The most significant finding does not merely concern the speed at which systems in the Claude family are evolving. According to researchers, as of October, AI itself had become the primary driver of the observed acceleration, with its contribution growing over time. Across the benchmarks examined, Anthropic reportedly packed about two years' worth of historical progress into a single year. This metric should be interpreted with caution, as benchmarks do not capture every dimension of a model's capabilities, but it remains an important signal: generative tools are increasingly being used to produce the research, engineering, and technical work required for the next generation of generative systems.
This does not mean a model has achieved full autonomy in designing, training, and deploying its successor. Human teams remain responsible for objectives, infrastructure, verification, and decisions. However, the gap between an assistant that speeds up individual tasks and a system that genuinely expands a lab's research capacity is substantial. If automation becomes useful across an increasing number of steps, the cadence of innovation can shift without waiting for a commensurate expansion of teams.
An acceleration not attributed solely to chips
For years, the trajectory of frontier AI has been told primarily through scale: more processors, larger datasets, more expensive training runs. METR says it is confident that the shift observed at Anthropic cannot be explained simply by an increase in compute capacity. The conclusion turns attention to how models are utilized within the organization: coding assistance, analysis, experimentation, and other research and development activities can produce compounding gains.
This distinction also matters from an economic perspective. Scaling up data centers requires capital, energy, supply contracts, and construction lead times; improving internal processes with software that is already available can, by contrast, make an impact much more quickly. The two dynamics are not necessarily mutually exclusive: the frontier continues to depend on highly expensive infrastructure. However, if models increasingly help make the research that improves them more efficient, a lab's competitive edge can no longer be measured solely in GPUs, data, and funding.
METR does not frame this evidence as proof of an imminent capability “explosion”. As described, the report captures a concrete acceleration rather than an end point. That is precisely what makes it relevant: debates around AI safety have long addressed the possibility that sufficiently capable systems could shorten the time needed to achieve further advances. Now, an investigation conducted inside one of the leading labs offers empirical evidence of an early version of that process.
The paradox for Anthropic and the industry
The publication comes at a sensitive time for Anthropic. The company is preparing for a potential stock market listing, while its CEO Dario Amodei has repeatedly called for labs to exercise greater caution in the development of advanced AI. The contrast is difficult to overlook: on the one hand, a call to slow down or put stronger safeguards in place; on the other, research that appears to find in AI a tool to advance faster.
This is not necessarily an irreconcilable contradiction. A company can argue that stronger capabilities are also needed to improve defenses, identify vulnerabilities, and develop evaluation techniques. But the METR report sharpens the practical question: who determines when acceleration becomes too fast to be understood, tested, and governed? The pace of development can outstrip that of safety evaluations, public policy, and even organizations' ability to remain accountable for their actions.
The case also interests Anthropic’s competitors. OpenAI, Google, and the other frontier AI leaders are investing in the possibility of using models to boost the productivity of researchers and developers. If the phenomenon identified by METR is confirmed more broadly, the competitive pressure to integrate AI into research cycles would become even stronger. In that context, a decision by a single player to voluntarily slow down could be difficult to maintain without shared agreements, external oversight, or rules applicable to the entire market.
The limit of transparency
METR’s evaluation offers a rare glimpse into the inner workings of a frontier lab, but it does not resolve the issue of public visibility. Under the agreement between the parties, Anthropic was permitted to redact passages from the final version of the document and reportedly made extensive use of this right. Consequently, readers, independent researchers, and regulators cannot fully reconstruct the data, methods, and considerations that led to the conclusions.
Confidentiality is not trivial: companies like Anthropic handle sensitive commercial information, technical details, and aspects that could carry security implications. However, when a report indicates that AI is making a decisive contribution to accelerating the development of more powerful models, omissions affect the ability to assess the scope of the phenomenon. Knowing that the effect exists is different from being able to understand which tasks it focuses on, what controls were used, how robust the results are, and where operational limitations still emerge.
The investigation is also separate from another program through which Anthropic granted security researchers employee-like access, launched following the summer hacking incidents. The distinction is significant: analyzing the security and exposure of systems is one thing; directly observing how the company uses AI to advance its own research is another. Both tracks are necessary, but they address different risks.
What to watch now
The next step will be to see whether Anthropic makes further non-sensitive details available and whether other labs agree to comparable evaluations. A single investigation is not enough to define a general rule for the industry, especially when portions of the material are confidential. It can, however, raise the bar: if companies claim to use AI to accelerate the work that produces more advanced AI, they will need to clarify more precisely which control thresholds they use and who can verify them.
For those designing rules and standards, the issue is not just the power of the model released to the public. The cycle that produced it also matters: how much time is needed to move from one generation to the next, what share of the work is automated, and whether risk assessments are keeping pace. The METR report does not provide all the answers, but it makes an issue that has often been treated as theoretical far more concrete. AI is not merely transforming products and professions outside the labs: it is also reshaping the industrial and scientific machine that builds it.
