Training Compute

METR, drawing on Anthropic's internal R&D data and researcher interviews, estimates the company is now improving its models at roughly twice its 2024-2025 pace, with about 80% of the acceleration attributed to Anthropic's own AI use rather than faster growth in training compute. The report also says Anthropic's public August risk report understated the speed, and finds no algorithm, compute or data bottleneck likely to slow the loop within 12 months.