Seven leading models rated on real knowledge work, with honest strengths, watchouts, and the cost-per-intelligence breakdown for each.