How do I know which AI model to use?
Research + my take
What the research says
Most AI companies sell their models in tiers. Anthropic’s Claude, for example, comes as Haiku (small and fast), Sonnet (the balanced middle), Opus (the most capable for complex work), and Fable (the newest top tier). OpenAI and Google offer similar small, medium, and large tiers under their own names.
The tradeoff is the same everywhere:
| Smaller models | Larger models | |
|---|---|---|
| Speed | Faster | Slower |
| Cost | Much cheaper per use | More expensive |
| Good at | Simple, repeatable tasks | Multi-step reasoning, nuance, long complex documents |
Public leaderboards and independent benchmarks (see the sources below) help you compare models. But they measure general ability, not your task. The most reliable test is to try two models on a real example of your own work and compare.
My take
I ask two questions: where does the work live, and how hard is the thinking?
Where the work lives. If the context is already in Outlook, Teams, Word, or PowerPoint (meeting recaps, email drafts, decks), Microsoft Copilot is often the right tool because it can already see that content. For long documents, contracts, comparing reports, or deep research, I reach for Claude.
How hard the thinking is. Here’s how I split the Claude models:
- Haiku: sentiment classification, bulk categorization, simple summaries, anything high-volume
- Sonnet: about 90% of my work, including SQL, Power BI, data modeling, writing, and real estate analysis
- Opus: important presentations, strategic thinking, investment analysis, complex business problems
- Fable: the hardest R&D-style problems
My rule of thumb: start with the middle tier. Step up when the answer isn’t good enough, and step down when the task is routine and you’re running it hundreds of times.