How do you choose between GPT-5's Auto, Fast and Thinking modes?

When GPT-5 arrived in ChatGPT on August 7, it came as one system rather than a list of models. Behind the scenes, a router (a small decision step) sends each message either to a quick chat model or to a slower reasoning model called GPT-5 Thinking. On August 12, after many users asked for more control, Sam Altman announced that paid users can now choose the behavior themselves from the model picker: Auto, Fast or Thinking.

The three options use the same GPT-5 family. What changes is the amount of reasoning, the hidden working-out the model does before it writes its answer. More reasoning usually means better results on hard problems and a longer wait; less means a near-instant reply that is fine for simple requests.

What each mode does

Auto is the default. OpenAI's help page explains that the system looks at your prompt and the conversation, and at patterns learned from how people choose models and how often answers turn out correct, to decide whether a request deserves deeper thought. When it does switch to reasoning, you see a short progress view and a Get a quick answer link if you would rather not wait.

Fast skips the extended reasoning and answers right away. It suits conversation, rewording, short explanations and anything where you can judge the result at a glance.

Thinking always reasons before answering. It takes longer, sometimes much longer, and is meant for multi-step problems. For Plus and Team subscribers, choosing Thinking by hand is capped at 3,000 messages a week, according to OpenAI's help page; times when Auto switches to Thinking on its own do not count toward that cap. Tom's Guide reports that the Thinking model handles up to 196,000 tokens of context (roughly the length of a long book), and that Altman said limits may change. Pro and Team users also see a Pro option that uses even more computing for the hardest questions.

Free users do not get the picker. They use GPT-5 with Auto routing, a limited number of messages every few hours, and one Thinking message per day before falling back to a smaller model.

A worked task table

The table below shows how one person might set the mode for a typical week. The choices are suggestions, not rules from OpenAI.

A simple way to read the table: if a mistake would be easy to spot and cheap to fix, speed wins. If a mistake could hide inside a correct-looking answer, give the model time to think.

Why Auto is a good default

Most of what people ask ChatGPT does not need extended reasoning, and waiting thirty seconds for a two-line reply feels slow. Auto aims to give you speed most of the time and depth when the request calls for it. OpenAI also says you can steer the router in plain words: phrases like "think hard about this" in your prompt count as a signal to reason more. That means you can stay on Auto and still ask for extra care on a single message.

Auto is not perfect. It sometimes treats a tricky question as easy, especially when the question is short. A request like "Is 1,001 a prime number?" looks simple but needs actual checking. If the answer matters, that is where your own judgment comes in.

When should you force Thinking?

Pick Thinking yourself when the task has a single correct answer that depends on several steps: math, scheduling with fixed times, comparing costs, reading a contract for a specific clause, or debugging code. In those cases a fast answer can sound fluent and still be wrong, and the extra reasoning gives the model a chance to check its own work.

Force it as well when Auto gave you a quick answer that you doubt. Instead of arguing in the same chat, open the picker, choose Thinking and ask the question again. If the two answers disagree, that disagreement is useful information: check the point yourself before you rely on either.

Finally, keep an eye on the weekly limit if you are on Plus. Save manual Thinking for work where accuracy matters, and let Auto and Fast handle the rest. For anything with consequences, such as health, money or legal questions, even the Thinking answer is a starting point to verify, not a final word.