7. Picking a Model, and Knowing When It Searched
The names change every few months. The two shapes behind them do not, and telling a live search from a confident recollection is the more useful skill.
The WJS Desk
Sep 16, 2026 · updated 8 days ago · 4 min read
There is a model dropdown at the top of the window. The names in it change every few months, the descriptions are short and vague, and picking wrong is either slow or worse. This last lesson is a way of choosing that survives the names changing.
Ignore the names, learn the two shapes
Whatever they are called this month, the menu offers two kinds of model.
| Fast | Thinking | |
|---|---|---|
| Answers in | A second or two | Ten seconds to a minute |
| Good at | Writing, rewriting, summarising, chat | Multi-step reasoning, analysis, maths |
| Usage cost | Light | Counts harder against your plan |
| Where it hurts | Problems with several dependent steps | Waiting for something trivial |
A useful default: stay on the fast one, and switch up when an answer disappoints in a specific way. Not when it is badly written, which the thinking model will not fix, but when it skipped a step, contradicted itself, or got arithmetic wrong.
Most people have this backwards. They set the most capable model permanently and wait ten seconds to have an email reworded, which buys nothing and burns the allowance they will want later.
Which to pick, by job
| Job | Pick |
|---|---|
| Write, rewrite, shorten, change tone | Fast |
| Summarise a document | Fast |
| Explain a concept | Fast |
| Anything with several dependent steps | Thinking |
| Comparing options against multiple criteria | Thinking |
| Arithmetic that matters | Thinking, and check it anyway |
The last row is deliberate. A thinking model is better at arithmetic, not reliable at it. In our Beginner track we caught a mortgage interest figure wrong by a factor of twelve, stated as flatly as everything around it. Model choice narrows that risk. It does not remove it.
Telling a search from a recollection
This is the more useful skill in this lesson, because it decides whether an answer is checkable.
The app can answer from what the model already knows, or go and look something up. Those two feel identical in the reply and are not remotely the same thing. Training data has a cutoff; a search happened just now.
Three tells that it actually searched:
- Citations or source links. The clearest signal by far.
- A visible searching step before the answer appears.
- It names something recent with specificity it could not have had.
When you need current information, say so rather than hoping:
Search for this and cite your sources. If you are
answering from training data rather than a search,
say so explicitly at the top.That second sentence is the one that does the work. It gives the model an honest exit, and a clearly labelled "this is from memory, it may be out of date" is far more useful than a confident answer you cannot place in time.
Where this bites hardest. Prices, version numbers, current officeholders, whether a company still exists, what a law says now. All of these change, all of them have a confident stale answer available, and none of them announce that the answer is two years old.
Two habits that beat picking correctly
Switch models mid-conversation rather than restarting. If a fast answer missed a step, change the model and say "redo that, you skipped the tax calculation." The conversation carries across, so you keep the context and only pay the slower model for the part that needed it.
Ask it what it is unsure about. One line, works on every model, and it is the highest value follow-up available:
What in that answer are you least confident about,
and which figures should I verify myself?It is usually right about the category even when it misses the specific instance, and the answer is a short list of things to check rather than a vague sense that you probably should.
A note on how fast this section dates
Everything else in this course is about mechanisms: permissions, file limits, where settings live. Those move slowly. Model names move every few months, and any article listing today's names by name is wrong within a quarter.
That is deliberate on our part. We have not named a single model in this lesson, because the two shapes and the search question outlive the menu. If the dropdown looks nothing like it did when you read this, the advice still holds: fast by default, thinking when a step was skipped, and always ask whether it actually looked anything up.
What model choice cannot fix
- A vague prompt. The most capable model still cannot know the facts you left out. Everything in our Beginner lesson on roles and reasons matters more than the dropdown.
- Accuracy on specifics. Numbers, dates, names and citations need checking on every model.
- Missing context. That is Custom Instructions and Projects from lesson 6.
The dropdown is the smallest lever in this course. Context is the largest, and it is the one nobody adjusts.
Where this leaves you
Seven lessons in, you can install the app and drive it from the keyboard, you know what the twelve permission prompts actually grant and that the app is not sandboxed, you can point it at a terminal or an editor instead of pasting, you know which voice feature is metered, you know the upload limits and which file types quietly fail, you have a setup that stops you re-explaining yourself, and you can tell a search from a recollection.
The honest summary is the same as everywhere else on this site: it is very good at language, structure and patience, and unreliable about specifics in a way that does not announce itself.
If you want the general skills rather than the app, the Super Beginner course starts from nothing, and the Beginner track after it is where the prompting techniques get tested rather than asserted.
