The timeline just got shorter. A lot shorter.
Last week, an independent research group called METR published results from a two-month evaluation inside Anthropic, one of the top AI labs in the world. METR (Model Evaluation and Threat Research) spent from August to October 2026 studying whether AI systems are now doing enough of the R&D work that creates future AI to actually accelerate that progress. The short answer is yes. The number they landed on is roughly 2x.
Anthropic’s models are now improving on well-defined tasks at about twice the rate they improved between March 2024 and October 2025. METR had access to internal Anthropic data that outside analysts rarely see, including training benchmarks, compute metrics, and direct interviews with engineers. What they ruled out quickly was raw computing power as the cause. No major hardware increase explained the acceleration. The driver is AI systems doing a substantial share of the R&D work itself.
Here’s what that looks like in practice. An AI researcher used to design experiments by hand, run them, interpret the results, and plan the next one. That cycle now happens with AI assistance at every step, compressing what used to take a year into something closer to six months. Roughly two years of historical research progress has occurred over the last twelve months at Anthropic.
METR puts real uncertainty on that 2x figure. The true acceleration could be anywhere from 1.3x to 3x the historical rate. They also did not identify obvious bottlenecks that would stop the progress over the next twelve months. No clear shortage of research directions to try, no data cliff, and no compute ceiling on the horizon. That absence of evident limits is what makes this finding worth paying close attention to.
The business implication is direct. The gap between what AI can do today and what it will do in eighteen months may be larger than the gap between 2024 and today. Organizations still running limited pilots instead of real deployments are optimizing for a capability curve that has already moved on. By the time a pilot report gets finalized and signed off, the tool it evaluated has changed.
None of this means every company needs to overhaul everything immediately. It does mean the cost of waiting is higher than it was a year ago. Getting familiar with AI in real workflows now, with real data and real feedback, builds institutional knowledge that cannot be acquired later in a weekend sprint. The organizations that will absorb the next wave of improvements are the ones already running the current generation.
Speed matters more than perfection here.
Want to explore how AI automation could benefit your business? Let’s talk.