Choosing a tier

The engine menu in the composer offers four choices, and the shortest useful answer is: leave it on Auto until a specific run gives you a reason not to. The other three are there for when you have that reason.

The menu itself is the definition. It reads, word for word:

Choice	What the menu says it is	Detail
Auto	Best value for each request	picks among the three below
Fast	Cheapest acceptable answer	The Fast tier
Smart	Balanced quality, long context	The Smart tier
Max	Deepest reasoning	The Max tier

What the tiers are actually dividing

Notice what those four lines are not about. They are not four price points, and they are not four speeds — despite everyone, us included, slipping into calling them that. Each tier is named for the kind of answer it is trying to buy — cheapest acceptable, balanced, deepest — and the price follows from that rather than defining it.

That distinction is load-bearing, because each of the three named tiers runs one particular model by default, and swapping which model a tier points at changes what that tier costs without changing what the tier is for. So a tier name on its own is never a price. Each tier’s own page — Fast, Smart, Max — names the model it runs and its rate together, and Model prices has the full table. Those pages are generated from the same source the product itself reads, which is why this page sends you there instead of repeating a number that would quietly drift.

Auto is the fourth thing, not a fourth level. It does not sit between Smart and Max. It reads the request you actually sent and picks among the three tiers for you — so what Auto costs depends on what you asked. A one-line question and a research job both sent on Auto will not land on the same tier and will not cost the same. We are not going to publish the rule it uses, partly because it changes, but the observable behaviour is the part you can rely on: Auto varies by request, and Auto is what every reply’s receipt records when you leave it alone.

Which job belongs where

Leave it on Auto for anything ordinary. Questions, drafting, tidying up a file, research you haven’t scoped yet. Auto is the default because for most of what people actually send, deliberately picking a tier is a decision that costs more attention than it saves money.

Reach for Fast when you already know the answer is easy and there is volume. Pulling fields out of a document you attached. Reformatting. Renaming forty things. Anything where you would be able to tell instantly that the output is wrong, so a cheaper attempt costs you nothing but the attempt.

Reach for Smart when the job is long rather than hard. Its menu line says long context and that is the signal: a big attachment, a chat that has been running for a while, a document you want read end to end and summarised faithfully. Smart is also what a handed-off task runs on when no engine was chosen for it.

Reach for Max when being wrong is expensive and you cannot check the answer yourself quickly. A decision that turns on reasoning rather than retrieval. Code you are going to run against something real. A document whose argument has to hold. Max is the one to choose deliberately and then stop choosing — compare the three tier pages’ rates before you leave everything on it, because standing on the dearest tier is how a week gets costly for no gain.

The thing tiers do not fix

A tier is a choice about how well the thinking is done. It does not change how much thinking a job needs, and on most bills the amount dominates the rate. A vague request sent on Max costs more than a precise request sent on Max, by a wider margin than Max costs more than Fast. If a run came back expensive, the tier is the second thing to look at — see Why it cost that much, and Writing a good request for the first thing.

Going below the tiers

Beneath the four tiers the menu offers Pick an exact engine, described as “Every model with real supply” — a list of every individual model you can run, by name, with sort and filter controls. That is the escape hatch for when you want a specific model rather than a category. Switching engines covers how to use it and what scope your choice applies to; Model prices is the full list with rates.

You can also point a tier at a different model permanently, so that Smart means something else for you from then on. Whether you should is its own question — see Changing what a tier runs.

Where to go deeper

This page is the overview: what the tiers mean and how to pick one. For a single tier’s specifics — the model it runs, its rate, a worked cost example, what happens when it cannot run, and what we have not verified about it — go to that tier’s own page: Fast · Smart · Max.