The AI evaluation firm METR has questioned the transparency of Anthropic’s public disclosures regarding the capability of its models to independently automate research and development. Following a review of the nonpublic version of the company’s August 2026 Risk Report, METR noted that while it agreed Anthropic had not yet crossed its internal acceleration threshold for AI self-development, the laity were given an understated view of how plausible such a threshold crossing was.
The finding follows a study that began in August to determine if artificial intelligence can drive its own technical progress using internal developer data. METR submitted a report of approximately 100 pages to Anthropic, but the evaluators stated that the company redacted the most informative pieces of evidence before the findings were shared with the public, forcing readers to rely on METR’s internal judgment of the results.
