MODEL TYPE
What Is a Reasoning Model?
A reasoning model is trained to work through a long run of intermediate steps before it gives a final answer, instead of replying straight away. That buys better accuracy on tasks with multi-step logic, and it costs more compute and latency per query.
What the video explains
A plain model, one token at a time
You have probably seen a chatbot say "thinking" before it answers. To see what it is really doing, the video compares two ways a model can answer the same question: a plain model on the left, and a model with thinking mode on the right.
A plain model answers one token at a time, and a token is a small piece of a word. To pick each token it runs one pass through the network, and every pass is the same size, so every token gets the same small slice of compute. A hard question gets no more effort than the word "hello," and the first thing it writes is final. That is fine for easy questions, but it struggles with multi-step math and logic, with long chains of instructions, and with catching its own mistakes, because it never looks back.
Thinking mode: scratch paper
Thinking mode adds one step. Before the final answer, the model writes out its work, step by step, in a reasoning span. Each token it writes is more computation, and the model reads what it wrote as it goes. It works like scratch paper.
This is still one single pass: the reasoning and the answer come out in one stream, with no second request. So the model can spot a bad step, cross it out, and try another way before it writes the answer.
The thinking budget
Most thinking models let you set a budget, which is how much scratch paper the model may use. Turn it low and the answer is faster and cheaper. Turn it to extra high and the model can think much longer, but it is slower and more expensive. You are buying accuracy with time and money.
A plain model goes straight to the answer, and a reasoning model works first, then answers. Use thinking mode when the problem has steps, and skip it when the answer is simple.
What's different
A standard model predicts the next token straight toward its answer. A reasoning model is trained, often with reinforcement learning on verifiable tasks like math and code, to first lay out a chain of steps and then answer from that chain.
This is not the same as telling a standard model to 'think step by step.' That prompt, chain-of-thought, nudges an existing model to show some working in its output. A reasoning model has the behavior trained in, with far longer and more thorough traces than a prompt alone produces.
What it costs
Reasoning tokens cost time and money. A reasoning model can burn far more tokens on the trace, hidden or visible, than on the answer itself, and that lands straight in your latency and your API bill. Providers usually price reasoning tokens the same as output tokens.
The payoff shows up on tasks with a checkable right answer and real multi-step structure: math, code, logic puzzles, planning. On casual conversation or a simple lookup, the extra deliberation adds cost and no accuracy.
When to reach for one
Reach for a reasoning model when a task has several dependent steps and one wrong step breaks the whole answer: multi-step math, debugging, planning a sequence of tool calls. Use a standard model for retrieval, formatting, summarization, and conversation, where the extra thinking changes the bill and not the outcome.
FAQ
- Is a reasoning model just a bigger model?
- No. Model size and reasoning capability are separate axes. A reasoning model can be built on a smaller base model and still outperform a larger non-reasoning model on tasks that need multi-step logic.
- Can I see the reasoning trace?
- It depends on the provider. Some return a summary of the trace. Others keep it hidden and hand back only the final answer, for product and safety reasons.