Why a new report on Anthropic automating AI development is alarming researchers
Anthropic is using AI to help build even more capable AI, and the results appear striking. The company’s models are improving at about twice the pace seen a year ago, with little sign that the trend is easing. A new report from METR, based on an 11-week examination inside the company, says AI is already speeding up Anthropic’s research and development in a major way, even as Anthropic continues to argue that AI progress should slow down.
The report suggests that much of the recent acceleration now comes from AI itself rather than simply larger training runs. METR found that Anthropic employees using its strongest internal model said they were about 12 times more productive on average, roughly double the level reported only a few months earlier. The researchers said they saw no clear bottlenecks or diminishing returns that would prevent even faster gains, and warned that fully autonomous contributions may soon play a much bigger role.
At the same time, the report highlighted how much remains hidden. Anthropic has completed several unreleased models, and METR said the company redacted much of the most informative material from the public version of the report. That has fueled concerns about transparency, especially because the company granted limited access rather than full visibility. Anthropic said the redactions were intended to avoid giving competitors a roadmap and to protect commercially sensitive material.
The findings have intensified criticism of Anthropic’s safety posture. Researchers and commentators say the report exposes a tension between the company’s public warnings about the risks of advanced AI and its active role in pushing the technology forward. Some argue the results show that AI-assisted self-improvement may be closer than expected, while others say the episode underlines the need for stronger independent oversight, more access for evaluators, and greater public transparency from frontier labs.
