Reasoning models like OpenAI's o1 and o3 work differently from standard AI: they spend extra computation time thinking through problems step-by-step before answering, which dramatically improves accuracy on complex tasks like competition math and coding, but comes with real trade-offs in speed and cost.

What Makes Reasoning Models Different From Regular AI?

A reasoning model is a large language model trained to work through a problem step by step before answering, rather than jumping straight to a final answer. The intermediate steps are called a "chain of thought." By spending extra computation generating those steps, the model handles multi-step math, code, and logic problems that ordinary models often get wrong.

The concept isn't entirely new. In 2022, researchers at Google discovered that simply asking any capable model to "think step by step" triggered intermediate reasoning that frequently fixed errors the one-shot answer would have made. However, the crucial shift since then has been from prompting for reasoning to training for it. Modern reasoning models like o1, o3, and DeepSeek R1 internalize this behavior during training, generating long, self-correcting chains automatically rather than relying on a user to ask them to think carefully.

How Do These Models Actually Learn to Reason Better?

The breakthrough behind o1, o3, and DeepSeek R1 is reinforcement learning applied to reasoning traces. The model generates step-by-step solutions, a reward signal scores whether the final answer is correct, and the learning algorithm pushes the model toward chains that lead to right answers. Over many iterations, the model learns to plan, check its work, and backtrack on its own.

DeepSeek's approach, detailed in a January 2025 paper, is particularly transparent. The company introduced DeepSeek-R1-Zero, "a model trained via large-scale reinforcement learning without supervised fine-tuning as a preliminary step," and reported that it "demonstrates remarkable reasoning capabilities," including spontaneously developing longer chains and self-verification behaviors purely from the reward signal. This matters because it shows that raw reinforcement learning can produce reasoning ability without human-labeled examples, which speeds up training and reduces costs.

What's the Real Performance Gain?

The accuracy improvements are substantial but depend on how much thinking time you allow. OpenAI made this explicit when announcing o1, stating that "the performance of o1 consistently improves with more reinforcement learning and with more time spent thinking". On the 2024 AIME math competition, o1 averaged 74% accuracy with a single sample per problem, rising to 83% with consensus voting over 64 samples and 93% when re-ranking 1,000 samples. This clean illustration shows that spending more inference compute buys more correctness.

Reasoning models excel where a problem decomposes into verifiable steps, such as competition math, algorithmic coding, scientific puzzles, and logical reasoning. They often outperform much larger conventional models on those tasks.

Steps to Decide When to Use a Reasoning Model

• Problem Complexity: Route hard, multi-step problems to a reasoning model and keep simple lookups, chat, and formatting on a faster, cheaper standard model.
• Verifiable Answers: Use reasoning models for domains with automatically checkable answers, such as math, competitive programming, and formal logic, where the reward function is cheap and reliable.
• Cost Tolerance: Understand that a single reasoning query can cost several times more and take seconds to minutes longer than a standard one, so budget accordingly.

What Are the Real Limitations?

The trade-offs are significant, and using reasoning models indiscriminately wastes money and time. The first cost is literal: generating long chains means many more output tokens, so a single reasoning query can cost several times more and take seconds to minutes longer than a standard one. The second limitation is "overthinking." Research has documented reasoning models burning hundreds of tokens to over-analyze trivial questions and even talking themselves out of correct answers.

Third, a longer chain is not a guaranteed-correct chain. The visible reasoning can look confident while reaching a wrong conclusion, a failure mode related to the broader problem of AI hallucinations. This means that even when a reasoning model shows its work, you can't assume the reasoning is sound just because it appears detailed and logical.

Reasoning models are slower, far more expensive per query, can "overthink" simple tasks, and longer chains do not guarantee correct ones. The practical guidance most teams converge on is to route hard, multi-step problems to a reasoning model and keep simple lookups, chat, and formatting on a faster, cheaper standard model.

Why Are AI Labs Investing So Heavily in This Approach?

The reason reasoning models have improved so quickly since late 2024 is that reinforcement learning is a good fit for reasoning in domains with automatically checkable answers. In math, competitive programming, and formal logic, the model can be trained on millions of attempts with no human in the loop for scoring. This is fundamentally different from preference-based alignment, where human raters must judge which answer is better. With reasoning models, a math answer is either right or wrong, no subjective judgment needed.

This objective reward signal has unlocked rapid progress. DeepSeek's production R1 model, which added a small amount of supervised data before reinforcement learning, "achieves performance comparable to OpenAI-o1-1217 on reasoning tasks". The fact that a company outside the US could match OpenAI's reasoning capabilities in a matter of weeks demonstrates how scalable and reproducible this training approach has become.