A model reportedly speeds internal AI research by up to 400x

Anthropic’s figure, relayed via an X post, is that Claude Mythos accelerated internal AI research by as much as 400x on tasks sized at about 40 hours of expert work. The same launch produced an apparent upward discontinuity on the Epoch Capabilities Index, recorded in the benchmark sweep, and Anthropic’s position, summarized by Eli Lifland, is that the resulting 2-4x steepening of the slope still falls short of the Responsible Scaling Policy trigger for a doubling of AI R&D speed. In the recursive-self-improvement trajectory it turns the qualitative claims of March, that 70-90% of Anthropic’s model code is written by Claude, into a lab-stated multiplier, one day after UNC’s 72-hour unattended run and eight days before Anthropic turned the same machinery on alignment research; the 34x kernel speedup found by Opus in February is the nearest earlier per-task figure.