Short answer
Use o1 when a task has one hard chain of logic to get right, such as a multi-step math problem, a tricky bug or a plan with many constraints. Stay with GPT-4o for everyday writing, quick questions, browsing, files and images, where o1 is slower, more limited and often no better.
Follow-up question What does the extra thinking cost?
Since September 12, ChatGPT Plus and Team subscribers have two new entries in the model picker: o1-preview and o1-mini. Both are trained to work through a problem privately before they reply, which makes them much stronger on some tasks and no better, or even worse, on others. The practical question is not which model is "smarter" but which one fits the job in front of you.
A good rule of thumb: if a careful human expert would need a pen, a scrap of paper and a few quiet minutes, try o1. If the expert could answer while chatting, GPT-4o is the better choice. GPT-4o is also the only one of the two that can search the web, read your uploaded files and look at images, so anything that depends on those stays with it for now.
What each model is built for
GPT-4o is OpenAI's general-purpose model. It answers fast, handles text, images and voice, and works with every ChatGPT tool. OpenAI itself says that for many common requests GPT-4o will remain the more capable option in the near term.
o1-preview is the first model in a new series that OpenAI numbers separately from GPT. Before it writes an answer, it generates a hidden chain of reasoning: it tries an approach, checks it, and backs up when something doesn't hold. On a qualifying test for the International Mathematics Olympiad, OpenAI reports that GPT-4o solved 13% of the problems and the reasoning model 83%. That gap is the whole argument for o1: it pays off where one wrong step ruins the result.
o1-mini is a smaller, faster reasoning model. OpenAI describes it as particularly good at code, and 80% cheaper than o1-preview, but it has less general world knowledge. Think of it as a strong problem solver who hasn't read as widely.
A worked comparison
Here is how the choice plays out on three common kinds of task. The judgments in the right-hand column are illustrative guidance based on what the models are designed for, not test results.
| Task | Example | Better first choice |
|---|---|---|
| Math with several steps | "A loan of $12,000 at 6% a year is repaid monthly over 4 years. Build the payment formula, then check it by computing the balance after month 1 and month 2." | o1-preview or o1-mini: the answer depends on every step being right, and the model checks its own work |
| Planning with constraints | "Schedule six volunteers across three shifts on Saturday. Ana can't work mornings, Ben and Chloé must not share a shift, each shift needs two people, nobody works twice." | o1-preview: many rules must hold at once, and a quick answer often breaks one |
| Everyday writing | "Write a friendly 120-word note inviting neighbours to a street clean-up." | GPT-4o: fast, natural, and extra reasoning adds nothing |
The writing row matters as much as the other two. Reasoning time does not make a short note warmer or clearer, and o1 can overthink simple requests. Send it the hard problems only.
How you write the prompt changes too. With GPT-4o, many people add "think step by step" to get careful answers. With o1, that step is built in, so a clear statement of the problem, the constraints and the format you want is enough. OpenAI's developer guidance also recommends giving o1 only the most relevant background material, because piling on extra documents can make its answer needlessly complicated.
Try it: Paste the volunteer-scheduling prompt from the table into GPT-4o, then into o1-preview. Check each schedule against every rule yourself, and note how long each model took.
Signs you picked the wrong model
You are probably using o1 where GPT-4o would do if the answer to a simple question takes many seconds to arrive, or reads as overly elaborate. You are probably using GPT-4o where o1 would help if you keep correcting the same logical slip, such as a schedule that breaks one rule every time, or a calculation that drifts by a few dollars. Switching models mid-project is fine: draft the plan with o1, then ask GPT-4o to turn it into a polished email.
What does the extra thinking cost?
The first cost is time. o1 does its reasoning before showing anything, and OpenAI's documentation warns that a request can take from a few seconds to several minutes, depending on how much reasoning the problem needs. For quick back-and-forth work, that wait adds up.
The second cost is your message allowance. At launch, Plus and Team users got 30 messages a week for o1-preview and 50 a week for o1-mini. In mid-September, OpenAI raised those limits: o1-preview now allows 50 messages a week and o1-mini 50 a day. A weekly cap of 50 is a good reason to save o1-preview for problems that deserve it and to try o1-mini first on coding and math.
For developers, the cost shows up on the bill. Through the API, which opened at launch only to accounts at the highest usage tier, the hidden reasoning is counted in "reasoning tokens." Tokens are the small chunks of text the API measures and charges by. You never see these reasoning tokens in the response, yet they are billed as output. OpenAI suggests leaving room for about 25,000 of them on hard prompts, so a short visible answer can still cost far more than the same answer from GPT-4o. o1-mini, at 80% below o1-preview's price, is the economical way to test whether reasoning helps your task at all.
Sources
- Introducing OpenAI o1-preview — OpenAI, September 12, 2024
- OpenAI boosts usage limits for o1 AI model — The Decoder, September 17, 2024
- Notes on OpenAI's new o1 chain-of-thought models — Simon Willison's Weblog, September 12, 2024