AI Tools Field GuideThinking modes and model tiers
Turn thinking on for hard, checkable work with a source attached, and leave it off for a rewrite.
What thinking mode does
A thinking (or reasoning) mode lets the model work through intermediate steps before it answers. Instead of writing the reply straight from the question, it first writes a working-out that only it reads, then answers from that.
The working-out is where it tries alternatives, notices a contradiction, and backtracks. It costs time and tokens, and it helps most on hard, multi-step problems. It does nothing for a rewrite, because there is nothing to work out.
| Turn thinking up | Leave it off |
|---|---|
| Multi-step reasoning about numbers, schedules with constraints (the arithmetic itself still goes to a tool) | Rewriting an email or memo |
| Multi-step logic: "if this, then what?" | Summarizing a document you attached |
| Code, data cleanup, anything a test will catch | Brainstorming and first drafts |
| Comparing options against several criteria | Simple questions with a source in the prompt |
| Anything where a wrong answer costs you | Bulk, repetitive tasks |
The left column has a pattern: the answer depends on holding several things in mind at once, and you can check the result. The right column is work where the model's first instinct is already good and deliberation only delays it.
Some products show a summary of the working-out. Read it as a narrative; it is not a proof. A model can show tidy steps to a wrong answer, and the steps shown are not always the steps taken.
What the names mean
Products label the same dial differently. Same idea each time: how much room the model gets to work before it answers.
| Label | Where you see it | What it means |
|---|---|---|
| Thinking tier or model | A separate entry in the model picker | The same generation with the working-out switched on |
| Reasoning effort low / medium / high | A setting beside the chat box | How long it may deliberate |
| Thinking budget | Developer and API settings | A cap in tokens on the working-out |
| Auto / Fast / Expert | Grok | Adaptive reasoning modes; Auto lets the model choose |
Some products expose no setting and decide for you; the advice on this page applies where you have the control. And watch the word "Pro": on one platform it is the bigger model, on another the deepest reasoning setting.
Model tiers: the other dial
Vendors ship each generation in sizes. The names vary (mini, flash, standard, pro, max) but the trade is the same.
| Tier | What you get | Use it for |
|---|---|---|
| Small / fast | Replies in seconds, cheapest, weakest on hard reasoning | Bulk reformatting, extraction from clean text, quick questions with the source attached, anything you will run a hundred times |
| Standard | The default balance | Most staff drafting, summaries, comparisons |
| Large / pro | Slowest, strongest on long documents and multi-step reasoning | The hard brief, the tangled instruction, the comparison that matters |
Tier and thinking are independent dials. Pick the tier by how hard the reading is, and the thinking by how hard the reasoning is.
| Thinking off | Thinking on | |
|---|---|---|
| Fast tier | Bulk reformatting, extraction | Rarely worth it |
| Large tier | Long document summaries: reads well, nothing to deliberate | The few tasks a month where being wrong costs more than waiting |
What it costs
Thinking multiplies the tokens a reply consumes, sometimes by ten, and the time with it. In a chat that is a longer wait. In an agent running many steps, or through a billed API, it separates a cheap job from an expensive one.
A staff example
| The job | Thinking off | Thinking on |
|---|---|---|
| A two-week watchbill: twelve people, three qualified for one watch, two on leave, nobody more than two nights running | A plausible draft that breaks the "two nights" rule twice | With the constraints as a numbered list, the model checks each in its working-out and the bill passes. "List every constraint and show it is satisfied" makes your check a one-minute job. Rewriting the watchbill memo afterwards is a thinking-off job. |
What to do with this
Find the thinking or mode control on the model you use most, and note the tiers on its model picker. Then read Reading a model's spec sheet so the next announcement makes sense.