So far reasoning models haven't been strong at, well, reasoning. If they are really like a killbot against a boxer now, that's pretty newsworthy