Thinking Mode

Programs like ChatGPT, Gemini, or Claude generate their answers word by word. Normally, such a program starts immediately with the first word of the answer. In thinking mode, it doesn’t. Instead, it first writes a longer text just for itself: it sorts out the task, tries intermediate steps, discards failed attempts. Only afterward does it formulate the answer that you get to see. You can think of it like scratch paper on a math test — the working-out done off to the side, before the result goes into the final fair copy.

What thinking adds in terms of accuracy

For simple questions, the mode brings little benefit. Anyone asking for the capital of Peru doesn’t need scratch paper. The difference shows up with multi-step tasks: mathematics, logic puzzles, programming, reviewing a long contract. There, models without intermediate steps often fail because they commit early and no longer notice the mistake. With an intermediate text, they can abandon a wrong path and try another.

Providers sometimes report large jumps for such tasks. On difficult mathematics test sets, the gains are double-digit percentage points, depending on the model. Precisely for this reason, thinking mode has become a selling point since 2024. By now, almost every major provider has its own variant of it.

The price for this is real. The thinking consists of additional text, and text costs computing time in the data center. An answer that would normally take two seconds can take half a minute instead and cost a multiple of that. For companies processing millions of requests per day, this is a serious cost factor. That’s why many providers offer a switch that lets you limit the length of the thinking.

What actually happens during thinking

Technically, thinking mode is not a new component in the model. It’s the same mechanism as usual: the model predicts the next word each time. The difference lies in the fact that part of these words is not intended for you. These intermediate words nevertheless end up in the text the model has in front of it. So they influence what comes afterward. The model uses its own intermediate text as a memory aid.

For this to work well, a model is specifically trained for it. It is made to solve many tasks for which the correct solution is known. Paths that lead to the correct result are reinforced; paths that miss the mark are weakened. This procedure is called reinforcement learning. After many runs, the model develops habits such as checking its own calculation.

A common misconception: that the intermediate text is an honest record of the internal processes. It is not. Studies show that models write down justifications that do not match the actual computational path taken. The text is a tool that improves the answer, not a window into the engine. Some providers therefore don’t display it in full at all, but only show a shortened summary.

The switch in chat windows and price lists

In most chat programs, you’ll find the mode as a selector above the input field or in the model list. Names and labels change quickly. Sometimes it’s called “Thinking,” sometimes “Reasoning,” sometimes it’s embedded in a model name with additions like “o3” or “Pro.” Often the software also switches on its own when it considers the question difficult. This becomes visible through a note like “thinking” and through the longer waiting time.

In business news, the term usually appears together with costs. Providers bill their prices by text chunks, and the intermediate steps count too. A thinking mode can make the same question a multiple more expensive. When a data center operator reports rising power demand, this is one of the reasons.

For your everyday use, a simple rule of thumb applies. For summaries, writing assistance, or quick follow-up questions, the mode is rarely worth it. For a physics problem, a bug in program code, or a decision with multiple conditions, it’s almost always worth it. And it doesn’t protect against false statements: a model can think carefully and still cite a made-up source.