METR: AI has roughly doubled the pace of model development at Anthropic

Everyone is used to hearing that AI writes code faster than a human. But there is a more consequential question: what happens when AI begins accelerating the development of AI itself? The independent organization METR tried to measure this at a real company. The result is intriguing and not as fantastical as one might expect, but it is far from dull.

What exactly was tested

Several METR employees worked inside Anthropic from August through October, with expanded access to internal data. Their task was twofold: determine how much faster model capabilities were improving and what share of that acceleration was specifically linked to the use of AI at work.

The main yardstick used to measure progress is called AECI. It is an internal version of Epoch AI's capabilities index. Since October 2025, it has been growing 1.9 times faster than before. Put simply, over the past year Anthropic has advanced about as much as it did over the previous two years.

What they found

The final estimate is that, by October 2026, the pace of model improvement had roughly doubled compared with the period from March 2024 through October 2025. The authors are candid that the figure is approximate and give a broad range of 1.3x to 3x.

The most interesting part is AI's own contribution. METR estimates that about 80% of the current acceleration comes specifically from using AI in research and development. In other words, most of the increase in speed comes neither from new GPUs nor from hiring, but from models helping to build the next generation of models.

Employees' subjective assessments are consistent with this finding. On average, they believe AI makes their work roughly twice as fast. In Slack surveys, individual estimates rose from 1.2x for Claude 3.7 Sonnet to 12x for the newest internal model. In the survey, 12x was equated with the statement: “I do in one day what used to take almost two weeks.”

What about the constraints?

One might expect growth to eventually hit a ceiling in algorithms, compute, or data. METR says it found no obvious constraints, at least over the coming year. None of the three potential bottlenecks is visible yet.

But the picture has weaknesses. First, the estimate was based on a single company and on internal metrics that cannot be independently verified from the outside. Second, the 1.3x–3x range is extremely broad: the lower bound looks like a meaningful improvement, while the upper bound looks like a different era. Third, employees' subjective estimates are always noisy, and the gap between the 2x average and the 12x peak shows that the effect depends heavily on the task and the individual.

What this means in practice

For developers and product teams, the takeaway is grounded. If the acceleration is only around 2x even inside the lab where people know the models better than anyone, it is risky for an ordinary team to plan around being “ten times faster.” It is more realistic to expect substantial but uneven gains.

For founders, something else matters more: the pace at which model generations turn over is likely to keep increasing. That means a product built around the limitations of today's model could become obsolete within months. A sensible strategy is to design an architecture in which the model can be replaced easily and to regularly reassess which workarounds are still necessary.

Finally, the methodology itself is worth watching. Independent audits of laboratories' internal data are becoming the norm, and reports like these—not marketing claims—will shape market expectations for the coming year.