How do I choose between all these models?
Picking from a menu that keeps changing.
Click the model picker and there's a list of names, version numbers and words like Fast, Pro and Max.
Ethan Mollick, a Wharton professor who writes a popular guide to this every few months, put it well: "The names do not map onto each other in any way that will help you remember them."
So here's how to choose without memorising any of them.
Most companies sell a family of models:
A bigger model doesn't know more trivia so much as it handles hard, messy, multi-step work better.
Mollick's advice for anything that matters: use "the most advanced models you can get access to", with thinking turned up (see "Making AI think harder").
For quick, simple jobs (reformatting, short rewrites, simple lookups), the smaller model is fine and much faster.
| Job | Model |
|---|---|
| Fix the formatting of this table | Small |
| Draft a reply to this email | Small or medium |
| Analyse this contract for risks | Largest, high thinking |
| Plan a data migration | Largest, high thinking |
Some apps have an "auto" mode that sends each question to whichever model it thinks fits.
When OpenAI launched GPT-5 in August 2025, this router broke on day one. Questions that needed the strong model went to a weaker one, and people concluded GPT-5 was dumber. Sam Altman admitted the problem and fixed it.
If quality seems to swing wildly, check whether you're on auto.
Epoch AI, a research group that tracks AI trends, found that the price of reaching a given level of performance has been falling fast, anywhere from 9 to 900 times per year depending on the task.
For performance like GPT-4's on PhD-level science questions, the price fell about 40 times a year.
So last year's "too expensive for that" is often this year's default. Re-check now and then.
Open-weight model: a model whose files are published, so anyone can download it and run it on their own computer.
Some are now good enough for real work. In 2025, Simon Willison ran OpenAI's open-weight model gpt-oss-20b on a MacBook Pro at about 82 tokens a second, which is fast enough to feel snappy.
The big benefit: nothing you type leaves your machine. That can matter for sensitive data.
Mollick's observation: the paid tiers mostly buy you more usage, and with agents that means more hours of the AI working for you, rather than a smarter answer to any single question.
If you only ask a few questions a day, a free tier using a top model may be enough. If you run agents all afternoon, you'll hit limits fast (see "Why your AI usage runs out so fast").
Default to the best model for anything important. Turn thinking up.
Drop to the small model for quick jobs. It's faster and saves your allowance.
Check whether you're on "auto". If quality swings, pick a model yourself.
For sensitive data, ask whether a local model is enough. Check with your IT team first.