Model names will change

Every week there’s a new model. We’re in a golden age of models, with enormous amounts of money going into training bigger and better ones. Names change, aliases move, limits get adjusted, prices drop, and older models get retired.

That’s why I don’t list a model ranking in this course. It would be wrong by the time you read it.

One thing is stable: the latest models are the best ones. Nobody uses an old model on purpose, except to save a bit of money. And since old models are retired regularly, even that doesn’t last.

So which model should I use? is not a question you answer once. It’s a dependency you manage, like a library version.

Write down what you picked

For a workflow that matters, record:

• provider and exact model identifier
• the date you evaluated it
• the prompt or system instructions
• relevant settings and available tools
• your evaluation examples and their scores
• the cost and latency you accepted

The model name alone isn’t enough. A changed prompt, a different tool description, or a new retrieval step can alter results as much as a model upgrade.

If the API offers a dated or pinned version, use it when you need repeatable behavior. A moving alias like latest is handy for experimenting. It’s also a way to have your application change behavior without a single line of code changing.

Treat an upgrade like a dependency upgrade

A newer model isn’t automatically better for your task. It might follow instructions better and still produce a different JSON shape. It might be faster but call tools more often. It might reject an input the old one handled fine.

Before switching, run your saved evaluation. Compare quality, latency, cost, and how it fails. If real users depend on it, roll out gradually.

Keep a way back. That can be the previous model identifier, a feature flag, or a routing rule that sends 10% of traffic to the new version first.

Keep the questions, drop the names

Names change. These questions don’t:

• Is the task simple or complex?
• How much context does it need, and what kind?
• Does speed matter?
• What does a wrong answer cost?
• Which tools or input types does it need?
• Can I evaluate the output reliably?

Your evaluation set is worth more than any list of this month’s winners. It lets you reconsider the choice every time the market moves, in an afternoon instead of a week.

For interactive coding, the honest answer is: try the current frontier models yourself, on your own project, and see which one gives you the best value. That takes an hour. Reading opinions about it takes longer and tells you less.