METR confirms Anthropic has doubled the pace of its R&D with help from its own AI, just as its CEO calls for slowing down

By The Seattle Times (AP) · November 5, 2026.

Anthropic is developing its AI systems roughly twice as fast as it was a year ago, according to an analysis released Thursday by Model Evaluation and Threat Research (METR), a Berkeley-based nonprofit. The finding comes as the company prepares to go public in what could be the largest IPO in history, expected this month, according to AP.

METR estimates that Anthropic is improving its models' capabilities on "well-specified tasks" at twice the pace of a year earlier. Its researchers attribute much of that acceleration to AI itself: Anthropic has said Claude is increasingly helping to build more advanced versions of itself.

The study draws on about 11 weeks of internal data, from mid-August to late October, and on interviews with employees. METR says it had "unprecedented access" to internal data. It also reviewed the risk report Anthropic published in August, along with a private version, and concludes that the public version "generally understates the level of acceleration" that was taking place.

Two technical caveats matter. First, METR says it can rule out "with confidence" that the cause is more compute (the computing power used for training). Second, it found no "bottlenecks," in either compute or data, holding back Anthropic's progress. The report itself acknowledges that AI does not yet make fully autonomous, large-scale contributions to the most critical parts of training, though Anthropic researchers describe "signs of life" that this could happen soon. METR also cautions that it may be overstating the effect, since the acceleration may be limited to those well-defined tasks.

The context sharpens the contrast. In a September 12 essay, CEO Dario Amodei called on the industry to "slow the pace at which we improve model capabilities" so that safety measures can catch up. In the same piece, he argued for placing external evaluators inside frontier labs and cited METR as a suitable example. OpenAI's Sam Altman said at the time that he also supported independent evaluators with employee-like access.

Sam McCandlish, Anthropic co-founder and head of responsible scaling, replied on X that many of METR's conclusions match his own: "AI is accelerating our research, but it's hard to say exactly how much." He acknowledged that the redactions in the report make it harder for others to "check METR's work," but justified withholding details on what drives the acceleration because it is "commercially sensitive" and because of the risk of "handing a roadmap" to competitors. According to AP, those redactions limited what METR could say about more detailed estimates.

**Our take.** The thesis: what matters is not the rhetorical contradiction, but that we finally have an external measurement of something that until now only the labs themselves reported. A third party with access to internal data confirming the effect—and adding that public communications understated it—sets a useful auditing precedent: checking what they say against what can be verified.

The Amodei paradox is less hypocritical than it looks, but not harmless either. Calling for a coordinated, industry-wide slowdown while you compete is a classic collective-action dilemma: if you slow down alone, you lose. Even so, an IPO incentive and a doubling of pace push in the opposite direction from the rhetoric. This is better read not as bad faith, but as evidence that voluntary governance has structural limits.

The most important data point is the one you can't see: with no visible bottlenecks and AI improving AI, the risk is that the loop compounds. For now it is an estimate with many caveats, limited to well-specified tasks, and the report's partial opacity prevents replication. Demonstrated capability must be distinguished from aspiration: fully autonomous AI contributions to its own training remain a signal, not a fact.

Overall, AI that accelerates its own development is also the path to what we advocate for in the long run: faster cures, abundance, work set free. But that horizon is only bright if the pace of capability doesn't decouple from the pace of oversight. In the short term, the sensible course is to demand more of this: independent evaluators, real access to data and fewer redactions. The winner is evidence-based governance; the loser, the "trust us" argument.