Models in 2026: fast answers vs thinking, and how to choose

0:00 Fast or thinking?

Some questions need an answer in two seconds. Others need a model that stops and works it through. ChatGPT can do both, and choosing well saves you time, money and embarrassing mistakes. In this lecture you will learn how ChatGPT's fast and thinking modes differ, what changed in twenty twenty six, and a simple rule for choosing every time.

0:26 Why this matters

Why does this choice matter? Because speed and depth really are a trade off. Pick a thinking model for a caption and you wait longer and burn through your allowance for no gain. Pick a fast answer for a pricing decision and you may get something that sounds sensible but skipped a constraint. Think of it like choosing between a quick phone call and a proper meeting. Both are useful. Using the wrong one for the job wastes time or produces a poor decision. So let's learn to choose deliberately.

1:05 Where answers come from

First, the mental model. Every answer comes from two places. What the model learned in training, and what is in this conversation right now, your messages, files, custom instructions, memory and the results of tools like web search, data analysis or connected apps. When an answer is wrong or generic, ask first, what was missing from the context? Very often, that is the real fix.

1:33 Two behaviours

Since the GPT five generation, ChatGPT combines two behaviours. Instant responses handle everyday questions, drafting, rewriting and summarising. Thinking responses reason for longer before answering, which is better for multi step logic, maths, analysis, coding and planning, but slower and more limited on usage. By default ChatGPT can route between them. On paid plans the model picker lets you choose, including thinking levels, and the most capable Pro models are reserved for higher tier plans.

2:06 2026: a fast-moving line-up

The line up moves quickly. Through twenty twenty six ChatGPT moved through several GPT five point x releases, and in September OpenAI announced its GPT six generation, including GPT six Astra, which it describes as its most intelligent model, and faster, lower cost models called Sol and Luna, rolling out to paid plans across the agent experience, Codex and the API. Names, defaults and limits will keep changing. So learn the pattern, fast, thinking and pro, and always check the picker and OpenAI's help centre.

2:43 The decision rule

Here is the rule. Rewriting, summarising, brainstorming, translating and quick questions, use Instant. Analysing a spreadsheet, comparing options, planning a budget or debugging a formula, use Thinking. High stakes analysis where depth matters more than speed, use the highest thinking level or a Pro model if your plan has one. And long, multi step work across apps and files, like building a deck from research, belongs to the agent experience we cover in module five.

3:16 Fix the prompt first

Before you upgrade the model, fix the prompt. Most weak answers are missing data, constraints or success criteria. A thinking model cannot invent your sales history. Give it the goal, the data, the constraints like total budget and minimum spend per channel, the success criteria, and ask for a table of workings, a list of assumptions and a sensitivity check, what happens if costs rise twenty percent?

3:45 Simple example

Here is a simple pair of examples. Rewrite this email subject line to sound friendlier. That is an Instant job. It returns in seconds, and three options are all you need. Now, which of three courier options is cheapest for forty orders a week, given different weight bands and a surcharge for remote areas? That is a Thinking job. The answer needs a small calculation table, and a fast guess could easily pick the wrong courier. Same person, same morning, two different right choices.

4:22 Worked example: a Dubai café

A café owner in Dubai asks what price to charge for a new breakfast set. In Instant mode, with no data, she gets a generic range. She switches to Thinking and provides ingredient costs, current prices, footfall by hour and a target margin. Now she gets a cost table, three price scenarios, a break even number of sets per day and a list of assumptions, including that labour is fixed. She corrects it, because weekends need an extra staff member, and the recommendation changes. The mode helped, but the data and her correction made it useful. Notice what changed and what did not. The model mode changed the depth of the working. The data she added changed whether the answer applied to her café at all. And the corrected assumption changed the final price. In her notes she now keeps three lines for every pricing question, the costs, the target margin and the staffing assumption, so the next analysis starts from the right inputs.

5:33 Pitfalls

Three pitfalls. Using the heaviest model for everything and hitting your limits by lunchtime. Believing thinking means verified. Reasoning models still make mistakes, especially with missing or wrong inputs, so check key numbers. And trusting model names from old tutorials. Success looks like this. Your default handles most tasks without rework, and you switch to thinking deliberately, then check the assumptions.

6:00 Usage and cost habits

A word on usage. Thinking and Pro models use more of your plan's allowance, and on the API they cost more because they generate more tokens. So draft and iterate in Instant, then switch to Thinking for the final analysis. Batch related questions into one well structured request instead of ten small ones. And on team plans, agree which tasks justify the heaviest models, so routine work does not burn through everyone's limits.

6:32 Try this now

Try this now. Pick one simple task and one complex task from your real work this week. Run each in Instant and in Thinking, in fresh chats with the same prompt and files. For each of the four answers, score the quality from one to five and note how long it took and how much you would need to edit. You will probably find Instant is enough for the simple task and Thinking clearly wins on the complex one. Keep the table, and re run it whenever OpenAI changes the model picker.

7:12 Watch me do it, part 1

Let me run the furniture store budget both ways. First I open the model picker and choose Instant. I paste the prompt from the lesson. Goal, a monthly ad budget split across Meta, TikTok and Google for a Lahore online furniture store in the fourth quarter. Data, six months of spend, revenue and cost per acquisition by channel. Constraints, a total budget, at least twenty percent on Google Search, a return on ad spend target of three. Output, a table, expected outcomes with assumptions, and a sensitivity check. The Instant answer arrives in seconds. It gives a sensible split, but no sensitivity check, and it quietly ignores the minimum Google share.

8:00 Watch me do it, part 2

Now I open a fresh chat, choose Thinking, and send exactly the same prompt with the same data. This time it takes longer, and the table respects every constraint, including the twenty percent Google minimum, and there is a sensitivity row for a twenty percent rise in cost per acquisition. I read the assumptions. Number three says costs stay flat in November, but November is our most competitive month. I correct it and ask for a rerun. The split shifts more budget to Google Search. In my model guide I record it, simple rewrites use Instant, budget planning uses Thinking.

8:43 Recap and next step

Recap. Answers come from training plus context. Use Instant for everyday tasks, Thinking for complex ones, and the highest levels only when stakes justify it. Fix prompts before upgrading models, and learn the pattern rather than the names. Your next step: run one simple and one complex task in both modes, score them, and keep the table as your personal model guide.