METR Says Anthropic's Model Development Pace Nearly Doubled
METR's 11-week assessment found that Anthropic's measured model-capability improvement rate since October 2025 was about 1.9 times the rate recorded from March 2024 to October 2025. Researchers estimated that internal AI use accounted for roughly 80% of the acceleration, while a separate employee poll suggesting 12x productivity was described as loose self-reported evidence. Anthropic redacted substantial evidence from the public report, and METR said the attribution estimate remains uncertain.
The findings provide an unusually detailed independent look at how AI assistance may be accelerating frontier-model research, while also carrying implications for evaluating recursive self-improvement and AI safety risks.
把员工自报的提效倍数,直接当成整个团队研发速度的倍数,是读这份报告最容易犯的错。
METR 在 11 月 5 日公布对 Anthropic 的评估,同一份材料既写到员工自报提效达到 12 倍,也写到模型能力指数的进步速度约为过去的 1.9 倍。
两个数字并不互相打架,因为它们没有测量同一个东西。
更容易误读的是,报告还讨论了进步加速有多少来自 AI,而这一层又不是直接测出来的事实。
下面把自报感受、指数趋势和因果判断拆开,再用一段能直接运行的代码检查其中的算术。
12 倍问的是员工感觉,不是团队产出
报告里的高倍数来自员工对自身提效的估计,并不是一组经过随机分配的对照实验。