Anthropic Says Claude Beat Best Listed Solutions on 72 MIPLIB Problems

The claim matters less for the raw score than for what it targets: MIPLIB is a reference benchmark, so an AI lab asserting it beat the best known listed solutions pushes models into territory usually occupied by specialized solvers. The caveat is built into the item — the tally is Anthropic's own and MIPLIB's public records haven't confirmed it, which makes independent verification the whole ballgame. Whether the submissions survive scrutiny, and how the benchmark community treats model-generated results, is the open question.