Anthropic's AI models appear to be improving at roughly twice their earlier pace on clearly defined tasks, according to an outside assessment published November 5. The researchers at METR, a nonprofit that evaluates advanced AI systems, found that Anthropic's use of AI in its own research likely explains much of the acceleration. They gave that judgment substantial uncertainty, and Anthropic withheld much of the internal evidence behind it from the public report.
METR measured how quickly successive models improved on a set of tasks. To judge what caused the change, its team also interviewed staff and examined internal data. The first measure tracks results. The second inquiry asks how much of the gain came from AI and what else might have contributed.
What the benchmark measured
The METR assessment compares the rate of improvement from March 2024 to October 2025 with the pace since then. On an internal version of the Epoch Capabilities Index, researchers estimated that scores were rising 1.9 times as fast in the later period. The statistical interval around that estimate ran from 1.5 to 2.8 times the previous pace. METR's broader judgment, which drew on additional information, was that capabilities on well-specified tasks were probably advancing at around double the earlier rate as of October 2026.
Think of 1.9 as a comparison between two slopes. At that pace, a year of progress on the index covered about as much ground as two years at the previous pace. It says little about how much better a particular Claude release will be at the work a customer gives it. It also says nothing about when a product will ship. The researchers warned that training focused on tasks similar to those in the benchmark, or quirks in the way the index works, could account for some of what they measured.
How AI helps the researchers
METR reviewed interviews, surveys and internal measures of model development. Staff described using AI to speed up research, although the size of the benefit varied. Some of the largest reported gains came from informal polls in which Anthropic employees estimated their own speedup. METR treated those answers as rough evidence. One person completing a task faster does not establish that an entire research program advanced at the same rate.
The team judged that AI probably accounted for most of the acceleration by October. It found that faster growth in training computing power did not explain the observed shift. Other factors may have played a larger role earlier in the period. Researchers said the work was still mainly AI-assisted and that the systems were not yet making highly significant, fully autonomous contributions to the most important parts of training. The assessment therefore describes people using increasingly capable tools to build new models.
What outsiders can check
Anthropic said in its announcement about independent evaluators that it had begun giving METR staff access comparable to that of employees. That arrangement gave the team a view of how research actually happens inside the company. It also allowed Anthropic to redact nonpublic information. As Edit USA previously reported, the terms let evaluators tell readers when important evidence has been withheld.
METR exercised that right in this report. It said Anthropic redacted most of the evidence the team considered most informative for its risk conclusions. Anthropic said the removals were aimed at keeping confidential information from accelerating other AI developers. Readers can examine the benchmark comparison and some evidence from interviews and surveys. Much of the material used to judge how long the speedup could last, and how much faster it might get, remains outside public view.
