Picture two friends answering a tricky math question.
- The first one blurts out an answer the second they hear it — sometimes right, sometimes a total guess.
- The second one pauses, grabs a napkin, works through the numbers, double-checks the math, then gives the final answer.
Today’s AI splits into these same two types — standard LLMs answer instantly, while reasoning models sit and think before they speak.
What is a reasoning model
Regular AI (ChatGPT-4o, standard Claude, standard Gemini) works by “thinking out loud as it goes” — the moment you hit send, it starts streaming out words one at a time, with no separate “thinking” step beforehand.
Reasoning models (OpenAI’s GPT-5.x Thinking models — the successors to o1/o3, DeepSeek R1, Claude with Extended Thinking mode, Gemini 3 with Thinking mode) are trained to do something called automatic chain-of-thought — before producing an answer, they generate a long internal stream of text you often don’t fully see: breaking the problem into steps, trying approaches, checking their own work, and only then writing the final answer.
The key point: this isn’t the same as typing “think step by step” into a prompt (which works on regular AI too). It’s a behavior baked into the model itself through a training process using reinforcement learning that rewards the model for reasoning correctly before it answers.
A concrete example
Try asking a regular AI this:
“Two trains move toward each other at 60 and 90 km/h, starting 300 km apart. When do they collide?”
A standard model can easily get this wrong, because it’s computing the answer “as it speaks,” one word at a time, with no chance to go back and check itself.
A reasoning model will spend a few seconds to tens of seconds “thinking” — setting up the equation, adding the combined speed, dividing by distance, converting units to minutes, double-checking the result — before writing a final answer that’s far more likely to be correct.
Open ChatGPT and turn the reasoning effort up (or hit “Think”), and you’ll see a “Thinking…” indicator or a short summary of the reasoning process before the actual answer appears — that’s reasoning in action.
When to actually use a reasoning model
Good fit — multi-step problems:
- Math, physics, statistics problems
- Writing or debugging complex code
- Planning with layered constraints (e.g. scheduling with lots of competing rules)
- Logic puzzles, probability, strategic analysis
- Reviewing contracts or legal documents that need cross-checking multiple clauses
Not needed (just slower and pricier for no benefit):
- Casual conversation, opinions
- Writing emails, social posts, short articles
- Translation, summarization
- Simple lookups that don’t require calculation
Simple test: if you ask a friend “what should I eat today,” nobody wants them to sit and think for 30 seconds before answering. But if you ask “help me calculate my taxes this year,” a bit of thinking time is well worth the wait.
The trade-offs
| Standard LLM | Reasoning model | |
|---|---|---|
| Speed | Near-instant | Slower (seconds to minutes) |
| Cost | Cheaper | More expensive (thinking tokens still cost money) |
| Accuracy on complex tasks | Prone to mistakes | Noticeably more accurate |
| Simple everyday tasks | Plenty good | Overkill — sometimes overthinks and gives a weirder answer |
| See the thinking? | No | Often shows a (partial) reasoning summary |
Something newcomers often miss: letting AI “think” isn’t free. Every word it reasons through — even the parts you never see — counts as billable tokens. If you’re using the API and leave reasoning mode on for everything, your bill can climb fast without you noticing.
There’s also a real quirk: reasoning models sometimes overthink simple questions, producing an overly complicated answer or burning time without any actual improvement — like someone taking way too long to work out 2+2.
How to tell you’re using one
- ChatGPT: turn the reasoning effort up to Medium/High/Extra High, or tap “Think,” on the GPT-5.x family (instead of leaving it on Instant)
- Claude: enable “Extended Thinking” mode in settings (available on newer Claude models)
- Gemini: select a “Thinking” variant under Gemini 3 or newer
- DeepSeek: use the R1 model, an open-weight reasoning model that’s free to use
If you see a “Thinking…” indicator or a box showing the reasoning process before the real answer lands, that’s your signal a reasoning model is at work.
Summary
A reasoning model isn’t “smarter” in every situation — it’s an AI trained to pause and think before it speaks. It shines on tasks with multiple steps or complex logic, but for everyday simple tasks, an instant-answer model is just as good and much faster.
Simple rule to remember: if the problem needs scratch paper, let the AI think first. If it’s just a chat, let it answer instantly.