Last month, an independent safety organization spent three months inside Anthropic running one of the most detailed audits ever conducted on AI development speed. The conclusion surprised even the researchers who ran it. AI capabilities are now improving at roughly twice the historical rate, and the main driver is not bigger hardware or larger datasets. AI itself is doing the work.
METR, a nonprofit that specializes in AI safety evaluation, was given access to Anthropic’s internal metrics from August through October 2026. They reviewed benchmark scores, training compute, and spoke with engineers across the company. Benchmarks are standardized tests that measure what a model can actually do, covering coding ability, reasoning, and problem-solving on a consistent scale. Those scores have been growing at 1.9 times the pre-October 2025 rate. Two years of progress, twelve months of time.
The mechanism matters. Anthropic did not buy new hardware.
What changed is that researchers are now using AI models to run the experiments that improve those same models. That is the loop. The more capable a model becomes, the more it can contribute to accelerating the research that builds its successor. METR searched for practical ceilings on algorithmic progress, training compute, and available data for the next twelve months, and did not find significant ones.
They were careful about uncertainty. The true rate could be as low as 1.3x or as high as 3x. Anthropic redacted portions of the assessment to protect competitive details, and METR was transparent about what was withheld and why.
For businesses planning around AI, the practical implication deserves real attention.
An eighteen-month AI adoption roadmap is a plan for a world that may not exist when you arrive. Capabilities that required a dedicated specialist team six months ago are showing up in general-purpose tools today. The gap between “this requires an expert” and “this comes with the software” has been closing fast, and the evidence now says that pace is itself increasing. Teams that build genuine understanding of what AI can and cannot do will adapt far more effectively than teams built around today’s specific tools and workflows.
The right response is not to rush. It is to plan with shorter loops. Run pilots that can expand when they work. Build in quarterly reviews of your AI roadmap. Treat it as a living strategy, not a document that gets written and filed.
The acceleration of AI is no longer a projection. It is a measured, documented result. Understanding that should inform how you structure your AI investments, your team, and your timeline.
Want to explore how AI acceleration could affect your business strategy? Let’s talk.