Anthropic’s AI development pace has doubled, outside group says

Anthropic’s AI capabilities for tasks with clear, measurable goals have been improving about twice as quickly as they did a year ago, according to an assessment released on November 5 by the nonprofit Model Evaluation and Threat Research (METR). The group examined company data and interviewed staff over roughly 11 weeks, from mid-August through late October, after Anthropic invited it to review its research and development process.

METR said Anthropic appeared to achieve benchmark gains over the last year comparable to those it had from October 2023 to October 2025. Its researchers believe AI systems helping develop newer models account for much of the acceleration. They said they could rule out increased computing power for training as the explanation, and identified no bottlenecks such as limits on compute or data.

The group cautioned that its estimate may overstate the pace of improvement. Its assessment focused on well-defined tasks that can be graded automatically or by a language model, and Anthropic may have understated its earlier progress. Redactions to the report also restricted the detail METR could publish, and its team had some internal uncertainty and disagreement.

Anthropic co-founder Sam McCandlish said many of METR’s conclusions matched his own and that the company would publish its own analysis in the coming weeks. He said the company withheld details about what is driving the acceleration because they could be commercially sensitive and help others advance faster.

The findings come as CEO Dario Amodei has urged the AI industry to slow improvements so safety measures can keep up. Anthropic is planning a public stock offering later in November. METR said the acceleration matters for safety because it affects how quickly risks from advanced AI may grow.

Anthropic’s accelerating AI raises urgent safety questions

Anthropic’s AI development is accelerating even as its CEO says the industry should slow down so safety measures can catch up. An assessment released on November 5 by the nonprofit Model Evaluation and Threat Research found the company’s capabilities for clearly defined tasks improving about twice as quickly as a year earlier.

The group reviewed internal data and interviewed employees over roughly 11 weeks. It said AI systems helping build newer models likely account for much of the faster progress, and researchers found no bottlenecks such as limits on computing power or data. METR warned that accelerating capability growth carries safety implications because it could make risks from advanced models grow faster, too.

There are caveats: the estimate focuses on tasks that can be readily graded, and the group said it could be overstating the acceleration. Redactions also restricted what it could disclose. But those limits underscore how difficult it is for outsiders to scrutinize progress when the key information sits inside AI companies.

Anthropic co-founder Sam McCandlish said the company would publish its own analysis in the coming weeks. He defended withholding details about what is driving the speedup as a way to avoid giving others a roadmap to faster development. Anthropic is preparing to go public later in November. The findings sharpen the case for independent evaluation as AI companies race ahead and debate how to manage the risks.

Anthropic’s AI gains show why progress needs clear-eyed scrutiny

An independent assessment puts a number on Anthropic’s rapid AI progress: its capabilities for well-defined tasks improved about twice as quickly as a year earlier. The finding, released November 5, deserves serious attention, not sweeping conclusions: the group itself warned its estimate could be too high and applies to tasks that are easy to grade.

Model Evaluation and Threat Research reviewed internal data and interviewed Anthropic staff over about 11 weeks. It believes AI itself is helping drive the faster development, while finding no limits on computing power or data that were holding progress back. That makes transparent, careful assessment essential as companies build increasingly capable systems.

Anthropic co-founder Sam McCandlish said many of METR’s conclusions matched his own and promised a company analysis in the coming weeks. He said the company withheld details about the speedup’s causes because they could be commercially sensitive and give competitors a roadmap. METR acknowledged that redactions make its work harder to check, a real cost when public understanding depends on access controlled by the companies being assessed.

CEO Dario Amodei has called for the industry to slow improvements so safety measures can keep pace. Anthropic is planning a public offering later in November. The debate should weigh both the risks of accelerating capabilities and the limits of what this assessment can establish, rather than treating one uncertain estimate as the whole picture.