Explained simply · Free to read · 5 min read

Reasoning model explained simply: when thinking mode in ChatGPT, Gemini and Claude is worth it

A reasoning model is an AI that thinks internally in several steps before answering: it breaks the task down, tries out approaches and checks them before it replies. In ChatGPT, Gemini and Claude you can choose this thinking level in the model menu. More thinking helps above all with calculations, planning and tricky analyses, but it takes longer and uses up more of your usage limit. For simple everyday questions, the quick answer is usually enough (as of October 2026).

Published 11 October 2026

In short

Reasoning model: A reasoning model is an AI that thinks internally in several steps before answering, breaking the task down and weighing up approaches. This helps with difficult tasks but takes longer and uses more computing power.^1,2,3,4; Example: You want to know whether a rail season ticket or single tickets work out cheaper for your family, and you enter your trips and prices. In thinking mode, the AI works through the options step by step; for asking when the first train leaves, the quick answer is enough.; Also called: thinking mode · thinking model · extended thinking · Deep Think · reasoning · Reasoning-Modell

What a reasoning model does differently

A normal language model writes its answer straight away, word by word. A reasoning model adds an intermediate step first. OpenAI describes it like this: the model generates internal thinking tokens, uses them to break the task down and weighs up different approaches before answering. Google says this thinking process improves reasoning and multi-step planning, for example in coding, maths and data analysis.

In practice you often see only a note such as "Thinking" and a timer. In Claude, Anthropic says you can open an expandable section that shows a summary of the thought process. That helps you follow how the AI reached its result.

Our assessment: thinking is a metaphor here. The AI computes for longer and writes down intermediate steps; this does not show understanding in the human sense.

Thinking mode in ChatGPT, Gemini and Claude

ChatGPT: the model menu has a slider for the thinking level with Instant, Medium, High, Extra High and Pro. According to OpenAI, Plus customers get Medium and High, and Pro customers additionally get Extra High and Pro. Free and Go do not get these thinking levels; there you can choose the Think option for harder questions, which uses a smaller model. Which one exactly is currently stated inconsistently by OpenAI: the model help page names GPT-5.6 Luna, while the German pricing table lists GPT-6 Luna as available for Free and Go and GPT-5.6 Luna as not available. Important: on Plus and Pro, Instant no longer switches to thinking by itself for hard questions, and asking it to think harder does not change the level either. Notebookcheck reported on this change on 17 September 2026 and gave a price of 23 euros a month for Plus.

Gemini: according to Google's help pages, you can adjust the thinking effort where the model supports it. Standard thinking is the default and answers faster. Extended thinking analyses the prompt for longer and is meant for complex problems. Deep Think is only available on the Google AI Ultra plan and only with the Pro model; an answer can then take several minutes.

Claude: Anthropic separates two settings. The effort level ranges from Low and Medium through High and Extra high to Max. On top of that there is a switch for thinking. On the newest models, including Sonnet 5.5 and Opus 5.5, Anthropic says thinking cannot be turned off.

When thinking mode is worth it

Anthropic lists typical uses as calculations and proofs, difficult coding tasks, comprehensive project planning, detailed document analysis and multi-step technical problems. For simple questions, general information and ordinary writing, Anthropic says the default settings are enough.

Google gives a similar rule of thumb: little thinking for simple factual questions, the default level for comparisons and creative reasoning, and the maximum for difficult coding, maths and multi-step planning.

Translated to everyday life: you do not need thinking mode to ask for a recipe or a translation. If, for example, you want mobile phone tariffs with several conditions worked out, or a travel plan with many dependencies drawn up, the higher level can be worth it.

The price of thinking: time, limits and cost

Thinking takes time. Google says standard thinking usually answers faster than extended thinking. Anthropic says higher effort levels give more thorough answers but take longer.

Thinking also uses more of your allowance. According to Google, higher thinking effort and more capable models use up your Gemini usage limit faster. One technical term helps here: tokens are small processing units for text, such as parts of words or punctuation; according to OpenAI, one token is roughly four characters of English text. Anthropic writes that higher levels use more tokens, so you reach your limit sooner. In ChatGPT, the limits for thinking levels depend on your plan; once the allowance is used up, OpenAI says ChatGPT may continue with another model.

For developers who pay per token, it gets directly more expensive: according to OpenAI, the internal thinking tokens, meaning the tokens for the intermediate steps, are not visible but are billed like answer text and so count towards usage. They also take up space in the context window. The context window limits how much content the model can take into account at once in a request. Google also bills all thinking tokens for Gemini.

Does thinking make the AI more reliable?

Not automatically. From its own test with difficult factual questions, OpenAI compares two models: gpt-5-thinking-mini admitted much more often that it did not know an answer, and was therefore wrong far less often than o4-mini. Both are reasoning models, though; OpenAI says o4-mini is also trained to think for longer before responding. So the comparison shows how differently models behave on errors and admitting uncertainty, not that switching on thinking mode by itself makes answers more reliable. That is a vendor figure.

An independent measurement shows that thinking is no guarantee. In Vectara's ranking of 22 September 2026, which measures how often models add made-up content when summarising texts, the thinking variants of some models even score slightly worse than their fast variants. Grok 4.1 Fast, for example, is at 19.2 per cent with thinking and 17.8 per cent without.

So even in thinking mode: check important details, figures and sources yourself.

Sources

Created with AI assistance and editorially approved before publication.