METR found that the pace of capability improvement for Anthropic AI models roughly doubled over the course of a year