---
title: Switching engines
description: Where to change which engine runs your work, exactly how far each of those places reaches, and when switching is worth doing at all.
---

There are four places to change the engine, and they do not have the same reach. Getting
that wrong is how people set something once, assume it stuck, and get surprised by a
receipt. Here is the whole map.

| Where you change it | What it applies to |
| --- | --- |
| The engine menu on the home composer | Your **account default** — every new chat from then on |
| Settings → *How it works* → **Engine** | The same account default. One value, two doors |
| The engine menu inside an open chat | **That chat only.** The account default is untouched |
| **Retry** on the newest reply | Re-runs that turn on the tier you pick — and that chat keeps it |

## How a chat gets its engine

A new chat takes a copy of your account default at the moment it is created. From then on
**the chat owns its own choice.** Changing it inside the chat does not change your default,
and changing your default afterwards does not reach back into chats that already exist.

So "switch the engine for one run" is really "switch the engine for this chat" — the choice
persists in that chat until you change it again. If you want a genuinely one-off engine,
start a new chat, set it there, and let it end with that chat.

There is no per-message engine picker. The smallest scope Else offers is the chat.

## Retrying a reply on a different engine

On the **newest** reply in a chat, the overflow menu offers **Retry** with a submenu of
**Auto**, **Fast**, **Smart** and **Max**. Picking one re-runs that same turn on that tier.
This is the fastest way to answer "would a better engine have got this right?" without
retyping anything.

Two things to know before you use it:

- **It re-runs the turn, so it costs again.** The original attempt was already charged; the
  retry is a second charge. See [Why it cost that much](/docs/why-it-cost-that-much).
- **It also switches the chat.** After a Retry on Max, that chat stays on Max for
  everything you send next. That is usually what people want and almost never what they
  expected, so it is worth saying twice.

Older replies do not offer Retry — only the newest one does, and only when nothing is
currently running. What older replies offer instead is **Branch in new chat**, which forks
a fresh chat carrying the history up to that reply. That is the tool for "go back to where
it was still on track and try a different direction" — see
[Reviewing what comes back](/docs/reviewing-what-comes-back).

## Picking an exact model instead of a tier

Under the four tiers, the engine menu has **Pick an exact engine**, labelled *"Every model
with real supply"*. That opens a list of every individual model Else can run, by name, with
sort controls including cheapest first.

The list shows model names, not prices. To read a name and its rate together, use
[Model prices](/docs/model-prices) — that page is generated from the same catalogue the picker
itself is built from, so it is always the current list and the current rates. What each model
can take *in* is sliced out of that same catalogue, a page per kind:
[models we measured reading images](/docs/models-that/read-your-images),
[models whose maker says they can read PDFs](/docs/models-that/read-pdfs), and
[models that take a whole document in one request](/docs/models-that/take-a-whole-document).

Two honest notes on exact picks:

- **What you pick is what runs.** Else does not quietly substitute a different model for
  the one you named.
- **Which is also the risk.** If the run then needs to use a tool and the model you picked
  cannot, the run stops rather than switching engines behind your back. For work that has
  to *do* things rather than just answer, a tier is the safer choice.

## When switching is actually worth it

Most of the time, it isn't — **Auto** exists so that this decision does not have to be
made per request, and it picks per request in a way a standing choice cannot. Switching
earns its keep in four situations:

- **You already know the job is easy and there is a lot of it.** Set the chat to Fast.
- **You already know being wrong is expensive.** Set the chat to Max, deliberately, for
  that chat.
- **A reply came back wrong in a way that looks like reasoning rather than missing
  context.** Retry it on a higher tier. If it came back wrong because it didn't know
  something, a higher tier will not fix that — give it the missing context instead. See
  [Writing a good request](/docs/writing-a-good-request).
- **You are comparing models on purpose.** Use Pick an exact engine, and read the
  receipts.

## Where the record is

Every reply in a chat carries a small receipt showing **the engine choice you sent it
with** — Auto, a tier name, or the exact model. That is the choice as you made it, recorded
before any routing happened, which is what makes it the reliable thing to check. See
[Reading a receipt](/docs/reading-a-receipt).

If you want to change what *Fast*, *Smart* or *Max* mean for you rather than which one you
use, that is a different control — see [Changing what a tier runs](/docs/changing-tier-defaults).