Anthropic's artificial intelligence models are improving on certain capability tests at an estimated rate 1.9 times higher since October 2025, according to an independent evaluation published by the METR organization on Thursday (5). The report, however, presents an important caveat: Anthropic prevented the disclosure of a large part of the evidence that the researchers considered most relevant to their conclusions.
The investigation was conducted between August 12 and October 29, 2026, with access to results from unreleased models, internal training metrics, research documents and interviews with company employees.
The estimate uses the AECI, a version adapted by Anthropic of the Epoch Capabilities Index, which tracks model performance on specific tasks. The result indicates that the company achieved approximately two years of historical progress on these tests during the past year. The 95% confidence interval ranges from 1.5 to 2.8 times the previous rate.
This acceleration does not mean that Anthropic's overall productivity has practically doubled. The indicator measures the speed of evolution of the capabilities evaluated, while AI's effective contribution to that progress depends on other evidence and interpretations.
Why part of the evidence cannot be publicly examined
Although it allowed an external evaluation, Anthropic retained the right to remove confidential information from the public report. According to METR, the company suppressed most of the evidence considered most informative for assessing the risks associated with accelerating AI development.
Among the restricted materials are detailed estimates of AI's own contribution to model development, data on algorithmic improvements, intermediate training indicators and projections about future systems.
Anthropic justified the cuts by the need to protect confidential information that could accelerate technological development at competitors.
METR acknowledged that the restrictions make independent evaluation of the arguments presented difficult. Because part of the evidence allows different interpretations, external researchers and readers must trust the organization's judgment to a greater extent than would be desirable.
The study also had access limitations. Although it received a significant volume of internal information, the team could not freely consult Anthropic's systems, and some data requests were denied.
The lack of transparency does not automatically invalidate the estimate of 1.9 times, but it prevents third parties from fully examining the elements used to support broader conclusions about the acceleration of research.
The organization says it found convincing evidence that the observed progress does not simply stem from faster growth in the computational capacity used in training.
Interviews and internal surveys also point to significant gains associated with the use of AI tools by researchers. However, these accounts are difficult to quantify, and METR found no decisive evidence that would allow all the acceleration to be attributed to artificial intelligence. Factors unrelated to AI may also have contributed, especially at the end of 2025.
Another important limit involves the autonomy of systems. According to the researchers, AI does not yet make fully autonomous and highly significant contributions in the most important stages of model training. Currently, progress occurs mainly with systems that assist human professionals.
The methodology itself requires caution. Tests with well-defined tasks do not necessarily represent performance on open real-world problems. In addition, analyses with different methodological choices produced acceleration estimates between 1.1 and 2.9 times.
METR considers it possible that more capable models will expand their contribution to the development of new generations of AI, producing additional acceleration. The report, however, does not demonstrate that this self-improvement process is already occurring autonomously and continuously.
