Extended Reasoning

Programs like ChatGPT generate text by appending, word by word, whichever next word fits best. Normally they start the answer right away. With Extended Reasoning, they don’t. Instead, they first write a kind of scratch note: intermediate calculations, intermediate considerations, discarded approaches. Only afterward does the actual answer for the user follow. The program isn’t really thinking, but it consumes more computational steps before committing to an answer — and that is precisely what makes it more reliable on hard tasks.

Why longer computation yields better answers

A program that answers immediately has the same amount of time for every question. For “What is the capital of Peru?” that’s enough. For a multi-step math problem, it isn’t. Anyone who solves such a problem in their head without intermediate steps makes mistakes. The very same applies here: the scratch note gives the program room to break the problem down into small steps.

For the AI industry, this was an important discovery. For a long time, the rule was: better answers come only from larger programs and more training data. Extended Reasoning shows a second path. You simply let the finished program work longer at run time. On math and coding tasks, success rates rise as a result, sometimes considerably, without changing anything about the program itself.

The price for this is real. Every scratch note consists of text that the program has to generate, and every generated word costs computing time in a data center. An answer with long thinking can be ten times as expensive as a fast one. That’s why Extended Reasoning is usually a deliberate choice among providers rather than the default.

The scratch note and how it comes about

Technically, the scratch note is nothing special: it’s text that the program writes for itself. It ends up in the area where the program keeps track of everything so far during a request. The program can thus reuse its own intermediate steps, much like you check a side calculation on your own scratch pad. Providers often hide this text or show only a summary.

For the program to produce useful notes, it is specifically trained to do so. It is made to work on tasks with known solutions, and the attempts that lead to the correct result are rewarded. Over millions of runs, it thereby learns useful habits: rephrasing the task first, trying out several approaches, checking the result again at the end. No one hand-writes these strategies in advance.

A common misconception is that the scratch note is a record of the actual thought processes. It isn’t. It is generated text that looks plausible and can nonetheless give a wrong justification for why the answer is correct. And more deliberation doesn’t help everywhere: for a simple factual question it accomplishes nothing, because missing knowledge isn’t created by mulling it over.

Reasoning models in products and headlines

In chat programs, you recognize the mode by toggles like “Think longer” or by a display such as “thinking for 24 seconds.” At OpenAI, the corresponding models are called o1 and o3; at Anthropic, there is “Extended Thinking” in Claude; at Google, similar variants of Gemini exist. The Chinese model DeepSeek R1 caused a stir in early 2025 because it offered this capability at low cost.

In business news, the topic often comes up in connection with data centers and chips. When programs compute longer per request, providers need more hardware for the same number of users. Analysts see this as one reason why demand for AI chips keeps rising. For corporate customers, the question is a practical one: which tasks justify the expensive mode, and which don’t?

Typical cases for Extended Reasoning are programming bugs, legal reviews, data analyses, and tasks with many conditions. For summaries, translations, or simple inquiries, the fast mode is sufficient. Many products now decide this themselves and switch on long thinking only for requests that are recognizably difficult.