
METR found Anthropic’s model capabilities on well-specified tasks improving at about twice the earlier rate. If AI shortens the development cycle, fixed review calendars can expire before their dates do.
The argument
The validity of a safety case depends on how far capabilities and operating assumptions have moved, not simply how many months have passed.
What is new
METR estimated 1.9× acceleration in Anthropic’s internal benchmark trend and judged that AI probably explains most of the current increase, while stressing substantial uncertainty.
Why it matters
A quarterly or annual approval can remain formally current while the model family, tools and development process have already outgrown its assumptions.
LecXie’s practical interpretation
Attach expiry triggers to model changes, new tools, faster R&D, autonomous contributions and disappearing bottlenecks. Reassess when evidence moves.
Evidence & limits
The estimate concerns well-specified tasks, not all capability. Much of METR’s most informative evidence was redacted, and fully autonomous AI R&D is not yet the main source of progress.
