AI Quick Tips 178: Reasoning models

Reasoning models are trained and tuned for tasks requiring, well, reasoning.  Not helpful, I know.  

They do well with tasks that require logic and/or problems that are complex.  In this case, complex means having multiple steps.

Most of the major AI chatbot tools have at least one reasoning model - ChatGPT, Deepseek, Gemini, etc.

The reasoning models break down problems into logical steps and then work through the steps one at a time.  Chain-of-thought, if you’re familiar with the term.  Some of the models will show you this process and wait for your approval, and some hide the process from you.  

Pros of reasoning models 

• They approach problems more like humans by breaking them down, working through them one step at a time, and exploring alternatives.
• They do better with logic problems than general-purpose models:
  • Math
  • Coding
  • Structured decision making
  • Etc
• They do better with complex problems.
• They sometimes give you a breakdown of how they are working through a problem.

Cons of reasoning models 

• They take longer to generate responses
• They cost more
  • They use extra “thinking” tokens that will get added to your output token costs.
• They use more resources
  • GPU
  • Memory
  • Etc
• They are less effective at smaller, simpler tasks - they “overthink” and can hallucinate more.
  • Single answer
  • Single step
  • Etc

Reasoning models do a really good job if they are given the right kind of problems to work on.  If you aren’t sure if it is the right kind of problem, then start with a general-purpose model, and if the answers aren’t good enough, try a reasoning model.