The research organisation METR has published an assessment of how artificial intelligence is affecting Anthropic’s model-development work. METR says its high-access review found that Anthropic’s internal model benchmark progress accelerated after October 2025, with an estimated 1.9-times rate compared with the earlier trend. The finding is important, but it is not a simple claim that every part of AI development has become twice as fast. METR describes substantial uncertainty, redacted evidence and a distinction between AI assistance and fully autonomous research.
METR compares research progress across two periods
METR says its assessment ran from 12 August to 29 October 2026 and involved access to internal information from Anthropic. The review included benchmark results for internal models, information about pretraining and reinforcement learning, compute use, algorithmic efficiency, AI-assisted research workflows and interviews with staff. METR describes the engagement as a high-access evaluation intended to estimate whether AI systems are accelerating the research and development of more capable AI systems.
The central question is consequential because faster AI research could shorten the time between model generations and make it harder for safety and governance processes to keep pace. METR says Anthropic provideed unusually broad access, but also notes that the public report cannot reveal all of the evidence it examined. Anthropic exercised redaction rights over a majority of the evidence METR considered most informative for its conclusions. That limitation affects how independently readers can assess the reasoning.
The 1.9-times estimate has important limits
METR reports that benchmark scores for Anthropic’s internal models, measured using the company’s AECI index, grew at an estimated 1.9-times the pre-October 2025 rate. The report gives a 95 per cent confidence interval of 1.5 to 2.8 times. METR explains this as roughly two years of historical benchmark progress occurring over the following year. The estimate is about the rate of change in a defined set of capability measurements, not a statement that every model capability or research activity doubled.
METR’s report says interviews and surveys with Anthropic staff were consistent with AI use contributing substantially to the faster pace of capability progress. Researchers described AI assistance as a major source of productivity, while METR says the evidence points more strongly to assistance than to fully autonomous contributions at present. That distinction matters. A model helping a researcher write code, analyse experiments or interpret results can speed up work without independently choosing and completing an entire research programme.
The assessment does not prove that AI systems can independently build the new of models, nor does it show that every bottleneck to AI development has disappeared. METR says it did not see evidence of significant bottlenecks in areSo as algorithmic progress, training compute, experiment compute or data over the following 12 months, but it also warns that its understanding changed as more information became available. A lack of observed bottlenecks in one review is not proof that no bottlenecks exist.
For AI developers, the report suggests that internal use of models may be becoming a meaningful productivity input to the research process. For safety teams and policymakers, the question is whether evaluation, interpretability, security and deployment controls can scale at a similar pace. If model capabilities improve more quickly because models help build successors, safety work may face a moving target. That possibility deserves attention even when the exact acceleration rate remains uncertain.
METR stresses that its estimate is uncertain and that Anthropic was allowed to redact much of the evidence it considered most informative. The report So provides an independent assessment based on privileged access, but not a fully reproducible public dataset. Readers should distinguish the measured benchmark trend from METR’s judgement about how much AI use caused it. The latter is a reasoned estimate, not a directly measured causal percentage.