That was definitely the case for Claude Opus 4.5, the most recent model of Anthropic’s strongest mannequin, which was launched in late November. In December, METR introduced that Opus 4.5 gave the impression to be able to independently finishing a job that will have taken a human about 5 hours—an unlimited enchancment over what even the exponential development would have predicted. One Anthropic security researcher tweeted that he would change the course of his analysis in mild of these outcomes; one other worker on the firm merely wrote, “mother come decide me up i’m scared.”
However the fact is extra difficult than these dramatic responses would recommend. For one factor, METR’s estimates of the skills of particular fashions include substantial error bars. As METR explicitly acknowledged on X, Opus 4.5 may have the ability to repeatedly full solely duties that take people about two hours, or it would succeed on duties that take people so long as 20 hours. Given the uncertainties intrinsic to the tactic, it was inconceivable to know for certain.
“There are a bunch of ways in which persons are studying an excessive amount of into the graph,” says Sydney Von Arx, a member of METR’s technical workers.
Extra essentially, the METR plot doesn’t measure AI skills writ massive, nor does it declare to. So as to construct the graph, METR checks the fashions totally on coding duties, evaluating the problem of every by measuring or estimating how lengthy it takes people to finish it—a metric that not everybody accepts. Claude Opus 4.5 may have the ability to full sure duties that take people 5 hours, however that doesn’t imply it’s wherever near changing a human employee.
METR was based to evaluate the dangers posed by frontier AI techniques. Although it’s best identified for the exponential development plot, it has additionally labored with AI corporations to judge their techniques in larger element and revealed a number of different unbiased analysis initiatives, together with a widely covered July 2025 study suggesting that AI coding assistants may really be slowing software program engineers down.
However the exponential plot has made METR’s status, and the group seems to have an advanced relationship with that graph’s usually breathless reception. In January, Thomas Kwa, one of many lead authors on the paper that launched it, wrote a blog post responding to some criticisms and making clear its limitations, and METR is at present engaged on a extra intensive FAQ doc. However Kwa isn’t optimistic that these efforts will meaningfully shift the discourse. “I believe the hype machine will mainly, no matter we do, simply strip out all of the caveats,” he says.
However, the METR group does assume that the plot has one thing significant to say concerning the trajectory of AI progress. “You must completely not tie your life to this graph,” says Von Arx. “But additionally,” she provides, “I wager that this development is gonna maintain.”
