Ajeya Cotra, a researcher at the evaluation group METR, argues in her newsletter Planned Obsolescence that Anthropic has probably already crossed the line it set for “dramatically accelerated progress due to AI.” She did not work on METR’s Anthropic report or see the unredacted version, and she says the opinions are hers alone.

AI Insiders has covered the two inputs. Anthropic’s IPO filing said it “may imminently meet” that threshold, and METR then reported that capabilities progress at the company ran about twice as fast over the year to October 2026 as in the twelve months before. Cotra’s contribution is the reading. Anthropic’s Responsible Scaling Policy sets the bar at twice the fastest sustained pace seen before AI helped with AI research, with the doubling plausibly caused by automation. On a plain reading, she writes, the bar has been met.

She concedes the edge case. METR’s authors credit AI with roughly 80 percent of the current speedup, which puts the central estimate against a no-AI baseline near 1.8 times as of their October writing. Cotra guesses it is higher now, and notes the policy also counts progress Anthropic merely expects.

Her own analysis starts from that doubling and works backward to the researchers. Model capability depends on compute and on software tricks that Anthropic calls compute multipliers, so for overall progress to double, those multipliers must improve roughly 2 to 3 times faster. Discovering them also takes experiment compute, which has not doubled. Under a deliberately simple model where labor and compute matter equally, she arrives at a 4 to 9 times speedup in researcher labor. Because easy ideas get used up as progress speeds along, she raises that to about 10 times today. Anthropic staff report a gain of roughly 12 times from internal models.

Cotra asks readers not to picture ten researchers where there was one. A better image is a time warp in which each researcher lives ten days for every outside day, reading, typing and thinking at ten times speed. She says no realistic round of hiring could have matched that.

The worrying part is what happens next. METR’s authors, she says, found no clear bottleneck and no steep diminishing returns that would stop the acceleration from compounding. If progress keeps accelerating smoothly, she calculates that the pace doubles again within six months, then in three, then in six weeks. In theory those intervals sum to a year. Real limits on software efficiency would cut that short, but she expects them to sit far from today’s models.

Her view of the stakes is stark. She expects systems that emerge from that loop could quickly develop novel weapons and beat humanity in a conflict, and that seriously misaligned ones would likely disempower humanity, possibly to the point of extinction. She points to Anthropic’s May 2025 policy, which required security against theft by state-level attackers at a milder milestone than this one. Cotra says that is the minimum, adds that the industry is nowhere near it, and notes that both OpenAI and Anthropic recently suffered major breaches by amateur groups. She also lists full monitoring, proven alignment methods, and governance tools for slowing down if safeguards fail.

Readers should weigh how much of this is her model rather than measurement. The 10 times figure is backed out of METR’s headline number using assumptions she calls very uncertain, and she says the true figure could sit well above or below it. The evidence underneath is mostly hidden: METR said Anthropic redacted a majority of what the authors found most informative. Cotra says she trusts her colleagues but wishes outside scientists could test their own models on the data.

Until that evidence is public, the only outside check on her ten-times estimate is the twelve-times figure from the staff whose work is being measured.

Ajeya Cotra, writing in Planned Obsolescence (planned-obsolescence.org); the post carries no publication date.