AI Is Now Writing Its Own Upgrades. What That Means for Your Business.
A few weeks ago, an independent research organization called METR spent three months inside Anthropic measuring something nobody had formally measured before: how fast is AI getting better at making itself better?
The answer should change how you think about your technology roadmap.
METR, which stands for Model Evaluation and Threat Research, is an independent nonprofit that evaluates frontier AI systems. This wasn’t a press release. It was a third-party audit with real access to Anthropic’s internal metrics, benchmark data, and interviews with their research staff. Anthropic is the company behind Claude, one of the leading AI assistants used in business today.
What METR found: since October 2025, Anthropic has been improving its models at roughly twice the rate it achieved between 2024 and 2025. Not because they hired more engineers or bought more computing power. The main driver is that AI is now doing a substantial portion of the R&D itself. Benchmarks, the standardized tests used to measure model performance, have been growing at 1.9 times the previous rate. Anthropic achieved roughly two years of historical progress in a single year.
That number is striking. But the mechanism behind it matters more than the number.
For most of computing history, software improved on a schedule engineers could estimate. Teams built roadmaps around it. If something wasn’t technically possible today, there were reasonable ways to project when it might be. That predictability is eroding. When AI systems start writing the code that trains the next version of AI, timelines compress in ways that are genuinely difficult to plan around.
The METR report is careful to note that most of this acceleration comes from AI assisting human researchers, not AI running fully on its own. That distinction matters for now. But researchers also noted clear signs that fully autonomous contributions will play a larger role soon. The trend is not slowing down.
For a business owner, this translates to two concrete realities. First, the AI tools you evaluated last year may already be significantly underperforming what’s available today. Pilots that felt underwhelming six months ago are worth revisiting. Second, the gap between companies actively building AI fluency and those watching from the sidelines is growing faster than it was. Each quarter of delay is a larger distance to close than the quarter before.
None of this requires a technical background to act on. The practical question is whether your organization is building the internal familiarity with AI workflows to keep pace with the rate of improvement, or whether you’re planning against a tool landscape that’s already obsolete.
The acceleration isn’t hypothetical. It’s been independently verified by a third party with full access to the data.
Want to explore how AI automation could benefit your business? Let’s talk.