Concepts

Reasoning models

4 min readintermediateUpdated 28 Sept 2026

1 · In one line

Reasoning models use intermediate reasoning tokens before producing a final response.

1 · What it is

Reasoning models generate intermediate reasoning tokens before they return a final response. Those tokens let the model break down a prompt, inspect alternatives and work through multiple steps.

OpenAI records reasoning-token counts in output-token details but does not expose the tokens through the API. DeepSeek returns reasoning_content separately from the final content. Google can return thought summaries. Anthropic supports a thinking-token budget on supported models.

Those tokens occupy context-window space and are billed as output tokens in OpenAI’s API. If the overall ceiling is too low, an API can end the request before a visible answer appears.

2 · Why it exists

Some tasks benefit from planning before the final response.

Multiple stepsHarder tasks may require the model to break down a prompt and consider more than one approach.

Tool workReasoning models can think between tool calls in supported OpenAI models.

Verifiable tasksDeepSeek-R1-Zero relied exclusively on reinforcement learning without supervised fine-tuning.

3 · How it works

Separate intermediate reasoning from the final answer.

OpenAI says reasoning tokens are used before a response. Google applies max_output_tokens to the combined total of thinking tokens and output tokens.

1 · readThe model receives the prompt.
2 · reasonIt spends reasoning tokens breaking down the prompt and considering approaches.
3 · actSupported models may reason between tool calls.
4 · answerThe model produces a separate final response.

Google returns only the final output by default.

4 · Where it's used

Who	What they ask	What it works with
Software engineer	“Which model should handle a multi-step debugging task?”	Reasoning support and effort controls
Research team	“How many reasoning tokens did this request use?”	Output token details
Agent developer	“Can the model think between tool calls?”	Interleaved thinking support

5 · What it solves, and what it doesn't

solves

Reasoning tokens give a model room to break down a prompt before answering.
Reasoning models can inspect alternatives during generation.
Some APIs let developers control reasoning effort or a thinking budget.

doesn't solve

Reasoning tokens still occupy context space and count as output tokens in OpenAI's API.
A low output ceiling can end a request before any visible answer appears.
Simple requests may need only minimal or low thinking.

6 · Go deeper

Chain of thought Inference Max tokens Reinforcement learning Tool use

Sources used

This explainer is written in original language. The links below support its factual claims.

docsReasoning models, OpenAI · read 28 Sept 2026
docsExtended thinking, Anthropic · read 28 Sept 2026
docsGemini thinking, Google AI for Developers · read 28 Sept 2026
paperDeepSeek-R1, DeepSeek-AI · read 28 Sept 2026
docsThinking mode, DeepSeek · read 28 Sept 2026

Next · 4 minMultimodal models A multimodal model can process more than one type of input, such as text, images, audio or video.

Suggest a correction