Three ways to see it

1. 1

   By now you know the landscape. You know the four major makers, the open-versus-closed split, the multimodal capabilities, and the rise of small local models. The remaining question is the one a colleague is going to put in your inbox by the end of the week: for this specific task, which model should we use. This lesson gives you a decision framework that will hold up across changing vendor pitches, across model releases, and across the kind of meetings where a senior officer needs to sound like they have actually thought about the trade-offs.

2. 2

   The six-factor framework. For any candidate task, ask these questions in order: what is the task type, drafting, reasoning, code, translation, classification; what is the context size needed, a paragraph or a hundred-page tender; what is the budget tolerance, free, low, premium, unlimited; what is the privacy requirement, public, internal, sensitive, classified; what is the latency requirement, real time conversation or batch overnight; what is the quality bar, good-enough first draft or publication-ready output. Six questions. Each one filters out a class of models. The model that survives all six is your answer.

3. 3

   A working cheat sheet for 2026, knowing it will need an update next year. For routine English drafting, Claude Haiku, GPT-5 mini, and Gemini Flash are all roughly equivalent and cheap. Pick the one your team already uses. For serious reasoning, multi-step legal analysis, board memos, complex strategic synthesis, Claude Sonnet or Opus is the safest current pick, with GPT-5 close behind. For code, Claude leads the public benchmarks for software engineering, with GPT-5 close behind, and Gemini behind both. For translation between English and Urdu in either direction, all three flagships are now comparable for general business prose, but you should test each one on your specific domain before standardising. For genuinely deep Urdu, particularly classical, poetic, or religious registers, all models still produce work that needs human editing.

The why-tree