METR Report Reveals Anthropic’s AI Acceleration: Models Improving Twice as Fast Amidst Redaction Concerns
METR’s new report reveals Anthropic’s AI models are accelerating twice as fast, driven by AI-assisted development, though significant redactions limit public insight into the full scope of risks.
S
··2 min readSciTech Editorial Team
Newsroom

A new report published by the AI safety research organization METR has revealed that Anthropic’s AI models are improving at a rate roughly twice as fast as they were a year ago. This finding, based on an 11-week internal investigation, suggests that Anthropic has achieved approximately two years’ worth of historical progress in just the last twelve months. The acceleration is alarming researchers because it indicates that AI is increasingly automating its own development cycle, raising urgent questions about the pace at which autonomous systems might emerge.
Crucially, METR researchers expressed confidence that this speed-up is not merely the result of scaling up computing power, which has historically been the primary driver of AI progress. Instead, the majority of the acceleration, particularly as of October, is attributed to the use of AI itself. Internal data, including informal Slack polls and employee interviews, showed that staff using Anthropic’s most capable internal models reported being around 12 times more productive than they would have been without AI assistance. This is a significant jump, roughly double the productivity gains reported just months earlier when using the Mythos 5 model.
The report highlights that much of this acceleration currently stems from AI assisting human researchers rather than operating independently. However, METR warned that there are emerging signs that fully autonomous contributions will play a larger role in the near future. Notably, the researchers found no clear indicators of diminishing returns or observable bottlenecks that could prevent even faster acceleration in the coming months. This lack of a visible ceiling is what makes the findings particularly concerning for the AI safety community.
Furthermore, the investigation revealed that Anthropic has been training advanced models at a swift pace. Over the past six months, the company completed training on three unreleased models, codenamed Model 2, Model 3, and Model 4, with training cycles finishing in July, August-September, and September-October, respectively. METR also identified an older unreleased model dating back to 2022. These internal models, which are more advanced than publicly available versions, are largely responsible for the productivity gains and acceleration described in the report.
Despite the transparency of the investigation, the report has been criticized for significant redactions. Anthropic was permitted to edit the final document and exercised this power extensively, removing a majority of the evidence that METR found most informative. The organization noted that it submitted a full 100-page report to Anthropic, but the public version lacks crucial details that could have multiple plausible interpretations. This has led to concerns that the public must rely heavily on METR’s judgment to assess the true risks and pace of frontier AI development, a situation that complicates independent scrutiny of one of the world’s leading AI labs.



