Anthropic says Claude agents beat known MIPLIB solutions and prove some optimal

Anthropic says Claude beat the listed solution to 72 of 203 open MIPLIB optimization problems and proved its solution optimal in 45 of those cases.

MIPLIB is a library of optimization problems such as airline crew scheduling and delivery truck routing.

In a separate claim, Anthropic says its agents found a timetable for 42 trains that beats MIPLIB's best known train scheduling solution by 2.6% in under two hours, and later proved that no better timetable exists. Anthropic says it has submitted the solutions to MIPLIB so others can verify them.