How to Use OpenAI: Pick the Plan, Set the Controls, Then Build a Project

How to use OpenAI in 2026 mostly means how to use ChatGPT well: pick the plan that matches how often you hit limits, turn off training on your data before you paste anything real, choose the thinking level per task instead of per mood, and put recurring work into projects. The developer API is a separate product with a separate bill, and it only earns its place once a task repeats.

That order matters. Most people start by arguing about which model is smartest, and they skip the two decisions that actually change their results: which plan they are on and where their work lives.

TL;DR

• Scale. OpenAI told TechCrunch on February 27, 2026 that ChatGPT had 900 million weekly active users and 50 million paying subscribers.
• Plans. Free is $0, Go is $8 a month in the US, Plus is $20, and Pro comes in $100 and $200 tiers. New Pro $200 sign-ups have been paused since September 10, 2026.
• The model line. Free and Go run on GPT-5.6 Luna. GPT-5.6 Sol starts at Plus. GPT-6 Astra reaches chat only as GPT-6 Pro on Pro, Business, and Enterprise, with 50 messages a week on Pro $100.
• Privacy. On personal plans, turn off Improve the model for everyone. On Business ($25 a seat monthly, $20 annually), OpenAI states it does not train on workspace data.
• Leave for the API when the same prompt runs more than about twenty times without a human reading every output.

What "using OpenAI" actually covers

OpenAI sells three things to three different people, and knowing which one you need saves you a lot of wasted setup.

ChatGPT is the product almost everyone means. It is chat plus tools: file uploads, image generation, voice, deep research, memory, projects, and scheduled tasks. ChatGPT Work and Codex sit inside the same subscription but have their own allowances, and they are where the longer-running and coding work happens. The OpenAI API is for software. It is billed per token, it does not come with any ChatGPT plan, and I already covered it end to end in the OpenAI API in 2026, so this post stays on the product side.

If you are not sure which bucket you are in, you are in the first one. Start there.

Step one: pick a plan by how often you hit limits

Every ChatGPT tier is a usage ladder with one or two capability jumps. The trick is knowing where the jumps are, because paying for more usage you never reach is the most common waste I see.

Here is what each rung actually buys, taken from OpenAI's own pricing page and help center.

Free. Unlimited everyday text chats on GPT-5.6 Luna, with limited uploads, image generation, voice, deep research, memory, and Codex. Enough to learn the product.

Go, $8. More of everything Free has, still on Luna. OpenAI's pricing page says the plan may include ads. Go makes sense if you are capped on uploads or images and never need heavier reasoning.

Plus, $20. This is the first real capability jump. OpenAI's model help page says Free and Go users do not get GPT-5.6 Sol at all, while Plus gets Sol at the Medium and High thinking levels. Plus also carries projects, scheduled tasks, and custom GPTs, and it includes GPT-6 Astra inside ChatGPT Work and Codex, though not in regular chat.

Pro, $100 or $200. Five times the usage, the Extra High and Pro thinking levels, and GPT-6 Pro (which is Astra) in chat. The allowances are finite: 50 GPT-6 Pro messages a week on Pro $100, shared with GPT-5.6 Sol Pro, and 200 a week on Pro $200. OpenAI's Pro tiers page confirms new sign-ups and upgrades to Pro $200 have been paused since September 10, 2026, so Pro $100 is the tier you can actually buy right now.

The arithmetic is short. Plus costs $240 a year. Pro $100 costs $1,200 a year, a $960 difference. If you bill $150 an hour, Pro pays for itself only if the extra usage and GPT-6 Pro save you more than about 6.4 hours a year, which is easy for someone hitting Plus caps every week and pointless for someone who hits them twice a year. Upgrade on evidence: the day the limit banner shows up twice in one week, move up a rung. Not before.

Step two: change one setting before you paste anything real

This is the step people skip, and it is the one I would do first.

On personal plans, your conversations can be used to train OpenAI's models unless you switch that off. OpenAI's data controls page spells out the path: Settings, then Data controls, then turn off Improve the model for everyone. When it is off, new conversations are not used for training, and they still appear in your history. Signed in, the choice follows your account across devices, and on a personal plan it also covers your Codex tasks.

For one-off sensitive work there is Temporary chat. Those chats skip your history, create no memories, are not used for training, and may be retained for up to 30 days for safety purposes.

If client or employee data is going through ChatGPT on a regular basis, the right answer is a workspace. OpenAI's Business overview states plainly that it will not train on your workspace's data. Standard seats cost $25 per user per month billed monthly or $20 billed annually, and Premium seats run $125 or $100 with five times the usage. A five-person team on annual Standard seats pays $1,200 a year, $300 less than paying month to month.

Step three: pick the thinking level per task

Since the GPT-5.6 rollout, ChatGPT gives paid users a thinking slider instead of a wall of model names. The ladder is Instant, Medium, High, Extra High, and Pro.

Two details from OpenAI's help page change how you should use it. First, on Plus and Pro, Instant does not automatically escalate when a request is hard. Second, typing "think harder" into the prompt does not switch the level. You choose it in the model menu or you do not get it.

My rule is to match the level to the cost of being wrong:

• Instant for anything you can verify in ten seconds: a rewrite, a formula, a quick summary of text you just read.
• Medium or High for anything with several steps or a number that matters: a pricing analysis, a contract read, a plan you will actually follow.
• Pro for work where an error costs you hours or money, and only when you have allowance left.

The same logic applies across every assistant, which is the argument I make in how to choose an AI model. The slider is cheap to move. Wrong answers delivered fast are not cheap at all.

The lineup moved again this week. On September 22, 2026, OpenAI announced GPT-6 Sol and GPT-6 Luna, rolling out first in ChatGPT Work and Codex for most paid accounts and in the API, with Luna coming to Free and Go users on desktop, according to TechCrunch. OpenAI's claim is that GPT-6 Sol makes about half as many mistakes as its predecessor. Treat that as a claim to test on your own work, not a reason to redo your setup this week.

Step four: put recurring work in projects

This is where ChatGPT stops being a search box and starts compounding.

A project is a folder with its own files, instructions, and chat history. Everything you ask inside it starts with that context already loaded, so you stop re-explaining your business every Monday. Projects, scheduled tasks, and custom GPTs are all on the Plus rung.

Here is how I set one up for a client:

1. One project per client or recurring job. Not one per topic. "Acme retainer" beats "marketing ideas."
2. A brief of about 200 words in the project instructions: who the client is, what good output looks like, the tone, and the three things never to do.
3. The files that answer the questions you keep asking: the price sheet, the brand guide, last quarter's report. Not the entire drive.
4. A standing first prompt you reuse. Mine starts with the deliverable, the audience, and the length, which is the structure I break down in how to write AI prompts.

Then use scheduled tasks for anything that repeats on a calendar: a Monday summary, a Friday checklist. If you catch yourself opening ChatGPT at the same time each week to type the same request, schedule it.

Step five: know when ChatGPT is the wrong door

ChatGPT is the workbench. It is not the factory.

Three signs you have outgrown it for a given job. You have pasted the same prompt more than about twenty times. The output goes somewhere other than your own eyes, like a CRM, a spreadsheet, or a customer. Or you need the task to run while you are asleep. At that point the prompt you refined in ChatGPT becomes code against the API, billed per token, and none of your subscription carries over. The OpenAI API guide walks through the first call, model pricing, rate limits, and structured outputs.

It is also worth checking whether OpenAI is the right vendor for the job at all. ChatGPT still has the widest tool surface, but for long writing or careful code review I often reach for something else, which I lay out in the best AI chatbot in 2026.

The bottom line

Using OpenAI well takes about twenty minutes of setup and then a weekly habit. Pick the plan by the limits you actually hit, which for most working people is Plus at $20. Turn off training before any real data goes in, or move to a Business workspace if clients are involved. Choose the thinking level by the cost of being wrong. Put every recurring job in its own project with a short brief and the right files. When a prompt starts repeating, move it to the API.

The people getting the most out of ChatGPT are not the ones on the most expensive plan. They are the ones whose Monday chat already knows the client.

Want the prompts I load into those projects? Grab Prompt Master, then join the newsletter for one practical AI build every week, tested before it lands in your inbox.