METR finds Anthropic’s model gains accelerated by internal AI usage

Anthropic’s large language models are improving at roughly twice the speed they did between March 2024 and October 2025, according to a report from the AI safety research nonprofit METR.reported The organization attributes about 80 percent of this acceleration to the company’s internal use of AI systems.reportedAnthropic is building better AI models much faster than before, largely by using AI to help build AI.METR data indicates an 80 percent contribution from internal AI workflows to Anthropic's doubled model improvement rate.

What it does

The report highlights a significant shift in how quickly Anthropic can iterate on its models. By leveraging AI tools within their own development workflows, the company has effectively doubled its pace of progress. This internal automation appears to be the primary driver behind the recent surge in capability gains.

How to get it

This information comes from a report published by METR. The findings were summarized in coverage by The Information, citing reporter Tiffany Li. There is no direct access to the internal tools described; these are proprietary workflows used within Anthropic’s own engineering teams.

Limits

The data covers the period from March 2024 through October 2025.reported The report does not detail specific technical architectures or code changes responsible for the speedup. It also does not provide granular metrics on which specific internal AI tasks contributed most to the acceleration.

The takeaway

Internal AI adoption is reshaping development velocity at major AI labs. For Anthropic, using AI to build AI has resulted in a measurable doubling of improvement rates. This suggests that organizational workflow changes may be as impactful as raw compute increases.

Building on this

Data science teams should evaluate how AI agents can automate parts of their own model evaluation and data curation pipelines. The documented acceleration at Anthropic implies that integrating AI into the development loop can significantly reduce iteration time. Teams might experiment with automated testing or synthetic data generation to mimic this efficiency gain. However, since the specific internal tools are not public, developers must adapt general AI assistance strategies rather than copying Anthropic’s exact stack. The key takeaway is that the bottleneck may be workflow, not just model size.