2-2 Prompt Basics: Non-reasoning Models vs. Reasoning Models

In this unit, we'll talk about when to choose reasoning models versus non-reasoning models when using Cursor.

What Are Reasoning Models?

First, let's talk about what reasoning models are. According to OpenAI's documentation, reasoning models think before answering questions. They generate a chain of thought internally, going through reasoning steps before giving their final response. These models are especially good at solving complex problems.

Examples of Non-Reasoning vs. Reasoning Models

The industry has many well-known non-reasoning models, like GPT-4.5, Claude 3.7 Sonnet, or DeepSeek's V3. The companies that make these models have also released corresponding reasoning models, such as OpenAI's o1 reasoning model and DeepSeek's r1 reasoning model.

To compare with a specific example, when you ask a question:

Non-reasoning models: Give you an answer right away based on the question - quick responses.
Reasoning models: The model first spends several seconds thinking, including planning and reasoning stages, before giving the final answer.

Choosing Models in Cursor

Back to our main topic - Cursor. In Cursor's settings file, you can see a model selection list that includes both the non-reasoning and reasoning models we just mentioned. For example:

Claude 3.7 Sonnet (non-reasoning mode)
Claude 3.7 Sonnet Thinking (reasoning mode)

This brings up an important question: When should you choose which type of model when actually using Cursor?

When to Use Non-Reasoning Models?

We recommend using non-reasoning models as your default choice in most situations - models like Claude 3.7 Sonnet or GPT-4o (the traditional models).

Here's why:

For simple tasks, non-reasoning models already perform excellently.
If you want quick answers or need AI to modify relatively simple code, non-reasoning models are more than enough.

For example, if you ask a reasoning model "What is this in JavaScript?" it might think for 19 seconds before responding, but the final answer won't be much different from what a non-reasoning model would give you immediately. For these clear, simple tasks, using a reasoning model is like "using a sledgehammer to crack a nut" - it wastes time.

As a result, for quick answers or simple code modifications, we recommend non-reasoning models for better efficiency without sacrificing quality.

When to Use Reasoning Models?

But there are situations where reasoning models work better, such as:

Complex multi-step tasks: Difficult bugs or features that need multi-stage planning.
When the problem isn't clearly defined: Reasoning models will think and plan before answering, and might even ask you questions to get more information, gradually clarifying the problem.

With traditional non-reasoning models, handling complex problems might require Chain of Thought prompts, but reasoning models already have this reasoning process built in from their training. This saves you from complicated prompt engineering. When facing difficult bugs or complex features, reasoning models often perform better.

Also, if your question is vague, non-reasoning models might give you a "hallucinated" answer, while reasoning models will break down the problem first and give more accurate responses.

Therefore, we recommend being flexible with model switching in Cursor based on your task needs - and experiment more:

Simple questions: Try non-reasoning models a few times. If you feel reasoning models take too much time, it might not be worth it, and you'll know to choose non-reasoning models for similar situations in the future.
Difficult problems: After trying non-reasoning models several times without success, switch to reasoning models. If they can solve the problem quickly and accurately, you'll understand when to choose reasoning models for similar situations.

After trying this approach several times, you'll gradually develop intuition for when to use which type of model.

Support ExplainThis

If you found this content valuable, please consider supporting our work with a one-time donation of whatever amount feels right to you through this Buy Me a Coffee page.

Creating in-depth technical content takes significant time. Your support helps us continue producing high-quality educational content accessible to everyone.