Anthropic’s AI models are getting better nearly twice as fast as they used to, and the main reason is the AI itself. A study published on Thursday by METR, an independent nonprofit that evaluates AI systems, found that the company’s benchmark performance has been improving 1.9 times faster since October 2025 than it did before. The researchers estimated that AI tools now account for perhaps 80 percent of that acceleration, as the systems help write code, run experiments and speed up the work of building their own successors.
METR, short for Model Evaluation and Threat Research, spent 11 weeks inside Anthropic between August and October. Its researchers reviewed confidential company research, interviewed staff and studied performance data from models that have not been released, according to Reuters and the Associated Press. The result is one of the clearest pieces of evidence yet that AI has started to accelerate its own development, a shift researchers call recursive self-improvement.
What the study actually measured
To track progress, METR used a score called AECI, Anthropic’s version of the Epoch Capabilities Index, which folds many benchmark results into a single number. Before October 2025, Anthropic’s score climbed about 15 points a year. Since then, it has risen about 29 points a year, which is how METR landed on the 1.9 times figure. In plain terms, Anthropic appears to have made about as much progress in the past year as it did in the two years from October 2023 to October 2025, the AP reported.
METR was also fairly sure about what did not cause the jump. “We are confident that the acceleration was not due to an acceleration in the scaling of training compute,” the group said, ruling out the simple explanation that Anthropic just threw far more computing power at the problem.
How AI speeds up the work of building AI
The speedup shows up in daily research work. Anthropic staff use AI to write and test code, design experiments and analyze results, and researchers at the company have reported sharp productivity gains, Reuters reported. Informal polls of employees suggested improvements of about 12 times on some tasks, though METR cautioned that those numbers were unreliable. Anthropic co-founder Sam McCandlish struck a similar note, saying that AI is speeding up the company’s research, but that it is hard to say exactly how much.
What AI still cannot do on its own
The study does not say machines have taken over the lab. METR cautioned that AI still cannot make major, fully independent contributions to the most important stages of training a new model. Humans still direct the research, set the goals and run the core training process. The acceleration the group measured also applies to what it calls well-specified tasks, problems with clear goals and clear tests, and METR said it could be overstating the effect beyond those tasks.
There are other reasons to read the numbers carefully. METR put high uncertainty on its own estimate, saying the true acceleration could plausibly be anywhere from 1.3 to 3 times the old rate. And because Anthropic reviewed the roughly 100 page report before release and removed a large share of the supporting evidence, outside experts cannot fully check the math, the AP reported. METR itself said Anthropic redacted a majority of the evidence it found most informative.
Why the finding worries safety experts
The bigger concern is pace. If models improve twice as fast, the tests used to check whether they are safe have to speed up too, and researchers have warned that advances could outpace efforts to assess the risks, Reuters reported. Recursive self-improvement is the scenario safety experts watch most closely, because a system that helps design its own successor could, in theory, improve faster than humans can monitor or control it.
The tension is sharpest inside Anthropic itself. Its chief executive, Dario Amodei, has publicly urged the AI industry to slow down so that safety measures can keep up, and he called for slowing advanced AI development as recently as September, according to Reuters and the AP. Yet his own company is now the clearest example of the speedup he warned about.
METR said it found no clear bottlenecks, in computing power, data or algorithms, that would force a slowdown over the next year. That means the race is likely to stay fast, and the open question for regulators and rival labs is whether safety evaluation can keep up. Follow the latest in our Technology coverage, including how jumpy investors have become about the cost of the AI boom in this week’s Nvidia selloff.