A practical, use-case-first comparison of the three big model families — no benchmark noise, just what to pick for coding, writing, analysis and agents.