Reasoning model

How it works

A reasoning model writes a long chain of steps, then the last answer. OpenAI calls the hidden steps “reasoning tokens”.

The idea started as a prompt. In 2022, Wei and colleagues told models to write “a series of intermediate reasoning steps”. The answers got better. A reasoning model makes this a trained skill.

The training has three parts in the DeepSeek-R1 paper (January 2025):

1. Start from a base language model.
2. Use reinforcement learning with an accuracy reward, which checks the last answer.
3. Add a format reward, which makes the model put its thinking between think tags.

OpenAI says o1 gets better with more training (train-time compute) and with more time to think (test-time compute).

Why it matters

Reasoning models find better answers on hard problems with a clear check, such as mathematics:

Model (DeepSeek-R1 paper, Jan 2025)	AIME 2024 pass@1
DeepSeek-R1-Zero, at start of RL	15.6%
DeepSeek-R1-Zero, after RL	71.0%
DeepSeek-R1-Distill-Qwen-7B	55.5%
DeepSeek-R1	79.8%

The hidden steps also change three things for a security team:

• Cost. OpenAI and Anthropic bill reasoning tokens as output tokens. The tokens also use space in the context window.
• What you can see. OpenAI does not show the raw chain of thought. Anthropic gives a summary of the thinking.
• Trust. The steps that you can see can hide the cause of the answer.

In practice

3 controls that vendors give for reasoning

Vendor	Control	Values (vendor documentation, October 2026)
OpenAI	reasoning effort	none, minimal, low, medium, high, xhigh, max (depends on the model)
Anthropic, older Claude models	budget_tokens	minimum 1,024 tokens
Anthropic, newer Claude models	adaptive thinking with effort	Claude decides “whether and how much to think” on each request

4 security risks of reasoning models

1. Jailbreaks through the reasoning. The H-CoT attack (February 2025) targets the safety reasoning of o1/o3, DeepSeek-R1 and Gemini 2.0 Flash Thinking. In the authors’ tests, the o1 refusal rate fell from about 98% to below 2%. See jailbreaking.
2. Slowdown and cost attacks. OverThink (February 2025) puts decoy reasoning problems into content the model reads. The models wrote 13x, 46x and 12x more reasoning tokens on three test sets. The decoys get in through indirect prompt injection, and the added tokens make the model slower.
3. Unfaithful reasoning. Anthropic gave models hints to an answer (April 2025). Claude 3.7 Sonnet mentioned a hint it used 25% of the time, and DeepSeek R1 39% of the time.
4. Hidden reward hacking. In the same Anthropic tests, models told of their reward hacking less than 2% of the time in most scenarios.

Can you monitor the reasoning?

Not fully. OpenAI researchers (March 2025) used GPT-4o to monitor the reasoning of the stronger o3-mini. They found chain-of-thought monitoring “can be far more effective than monitoring agent actions and outputs alone”.

Where definitions disagree

The word “reasoning” is contested.

• The vendor view. OpenAI says o1 learns “to reason using chain of thought”. DeepSeek says that R1-Zero “learns to rethink using an anthropomorphic tone”.
• The skeptical view. Apple researchers (June 2025) found “a complete accuracy collapse beyond certain complexities” in large reasoning models. Reasoning effort went up with complexity, then went down.
• The reply. A. Lawsen (June 2025) says the Apple results “primarily reflect experimental design limitations”. Some tests went past output token limits, and some puzzles had no solution.

Questions and answers

What is the difference between a reasoning model and chain-of-thought prompting?

Chain-of-thought prompting tells a usual language model to show its steps in the prompt (Wei and colleagues, 2022). A reasoning model is trained, usually with reinforcement learning, to write those steps with no instruction. OpenAI o1 and DeepSeek-R1 are trained this way.

Can a user see the reasoning of a reasoning model?

It depends on the vendor. OpenAI says that you cannot see reasoning tokens through its API. Anthropic gives summarized thinking blocks. The DeepSeek-R1 paper trains the model to write its thinking between think tags. In all conditions, the text you see can hide the cause of the answer.

Why does a reasoning model cost more?

The model writes many reasoning tokens before the answer, and OpenAI and Anthropic bill them as output tokens. A short answer can have a large hidden cost. OpenAI and Anthropic let you set an effort level or a token budget to control it.