Which GPT Model Should You Actually Use?

Every few months the model list changes. A new name appears, an old one retires, and the question starts again: which one should I be using?

It is the wrong question. “Best” is not a property a model has on its own — it depends entirely on the task in front of you. A more useful question is much simpler to answer, and it keeps working after the menu changes.

Don’t ask which model is smartest. Ask which model is sufficient.

Default state · one visible choice

Advanced controls revealed · Model + Effort

By default one selector is visible; the advanced controls reveal two separate choices — Model and Effort. Interfaces change, so treat this as what it looked like when captured, not a permanent layout.

ChatGPT desktop interface captured August 2026 · light mode

The first decision is capability: which model answers. The second is reasoning effort: how hard it thinks before it answers. Reaching for a more capable model and asking for more reasoning are different moves, and confusing them is why so much model advice feels contradictory.

Choose capability

Models are optimised for different kinds of work. They are alternatives, not a ladder — the fast one is not a worse version of the careful one.

Choose reasoning effort

Effort is about how much thinking the task deserves. Escalate when the work is complex, ambiguous, or built on several connected steps — not simply because the work feels important. And a short question is not automatically a low-effort one: fresh factual information may still need retrieval or search, whichever effort level you pick.

Reasoning-effort selector

Five named levels, in one menu. The names may change; the decision they represent does not.

ChatGPT interface captured August 2026 · light mode

Two things worth knowing about the top of the range. More reasoning is not a substitute for verification — the higher the consequence, the more review the output needs. And Pro is not simply “Extra High, but more”: Medium, High and Extra High vary reasoning effort on the same complex-work model, while Pro uses a higher-capability option intended for especially difficult tasks and longer-running workflows.

The framework, working

Task A — tidy up a customer email. One short piece of writing, you already know what you want, and a mistake costs a second read. Low complexity, low consequence: a fast general-purpose model at Instant effort is sufficient. Reaching higher adds waiting, not quality.

Task B — evaluate three suppliers against six constraints. Several moving parts, trade-offs that interact, and a decision the business will act on. The reasoning is the deliverable, so a model built for complex professional work at High effort is justified — and you still read the output critically before acting on it.

The lesson is not “use the strongest model for important work.” It is that capability and reasoning should be justified by the task and by the cost of getting it wrong.

What changes the economics

Four things move the answer, whatever the model names happen to be:

• Frequency and volume — a choice you make once is not the same as a choice made ten thousand times a day.
• Latency sensitivity — more reasoning takes longer, and sometimes the wait is the whole problem.
• Cost — capability and reasoning are never free, they are just charged differently.
• Consequence — the cost of being wrong is what justifies spending more capability, and what raises the need for human review.

Those last two behave very differently depending on where you are working. For an individual using ChatGPT, the economics show up as a subscription and usage limits: the practical cost of choosing more effort is mostly your own time waiting. For a system calling an API, every request is metered, reasoning consumes additional tokens, and a choice that is trivially affordable once becomes a real budget line at scale.

Prices, limits, model names and credit systems all change. Keep the mental model, not the numbers.

The product isn’t the model

One more distinction worth holding: the product you are using — ChatGPT, Microsoft Copilot, GitHub Copilot, Codex — is not the same thing as the model underneath it. Different products expose different models, tools, controls and configurations. Two people can be working with the same model and still get different results, because the instructions, context, tools, memory and permissions around it differ.

That is why the product is a step in the sequence, not a footnote: it decides which capability and effort controls you actually have to choose from.