The Model Landscape: Which to Choose for What

There are a lot of models, and new ones come out every month. The good news: you don't need to know them all or chase the 'latest and greatest.' You need a map to choose wisely, one that will still serve you even if three new models come out tomorrow. Here's the one I use.

The Big Ones (Closed, Powerful)

These are the most capable models right now. 'Closed' means you use them through the company that made them (you don't run them yourself); in return, you get top quality without having to manage anything. These are the three that matter today:

• Claude (Anthropic): excellent for code, long-form reasoning, and carefully following instructions. My favorite for building.
• GPT (OpenAI): very versatile, good ecosystem, strong on general tasks.
• Gemini (Google): strong in multimodal (image/video) and with a huge context window; well-integrated into Google Cloud.

The Open Ones (Open-Source)

Models like Llama or those from Mistral can be run by you on your own servers. The upside: control, privacy, and cost at scale (no one sees your data and, with volume, it's cheaper). The downside: more setup work, and they're usually a step behind the big ones in capability. They're great for specific use cases, not for your first product.

The 4 Knobs You Decide on Every Time

Here's the map that doesn't expire. Choosing a model isn't about memorizing names; it's about turning four knobs depending on the task. No single setting is 'the right one' all the time: it depends on what you're doing.

• Capability — how smart/precise you need it to be (reasoning, difficult code).
• Speed — how fast you need the answer (is it interactive? is it in bulk?).
• Cost — how much each answer costs you (matters a lot when you're talking millions).
• Privacy — how sensitive the data is (can it leave your servers or not?).

Worked Example: Two Tasks, Two Models

Task A: writing the complex logic at the core of your app. Here you want MAXIMUM capability; speed and cost matter very little because you only do it a few times and you need it to be right. Capability knob to the max → the most powerful closed model you can afford (for me, that's Claude).

Task B: classifying 100,000 user comments as 'positive/negative.' It's simple but massive. Here, top capability is an expensive waste; what matters is cost and speed. Cost and speed knobs front and center → a cheaper, faster model. Same person, same week, two different models — because the task is in charge, not the brand.

The Rule That Saves You Time

Don't marry one model. The 'right' model depends on the task: capability vs. speed vs. cost vs. privacy. Learn to turn those four knobs, and you'll always choose well, even with new models that don't even exist today. The map outlasts the models.

Your Takeaways

• You don't need to know every model; you need a framework for choosing.
• Large/closed = maximum capability with zero management (Claude, GPT, Gemini).
• Open-source = control, privacy, and cost at scale, in exchange for more work.
• You choose by turning 4 knobs: capability, speed, cost, and privacy.
• The task decides the model, not the hype or the brand name.

The best model isn't the newest one this month: it's the one that fits the task you have in front of you today.