Logos52
wiki / Systems / AI and Agentic Systems / Thinking Models

Thinking Models

concept updated 2026-06-15

Thinking Models

Some LLMs improve difficult-task performance by spending more internal computation on problems such as math, coding, and logic.

Summary

Karpathy’s practical advice is to use fast models by default, then switch to a thinking/reasoning model when the problem is hard or the first answer needs deeper work.

When To Use

Use thinking models for:

  • Debugging difficult code.
  • Math and logic.
  • Multi-step reasoning.
  • Ambiguous technical diagnosis.
  • High-value decisions where latency is acceptable.

Avoid them for:

  • Simple recall.
  • Basic ideation.
  • Low-stakes chat.
  • Tasks where extra latency adds no value.

Always Want the Smartest Model

Naval’s selection rule runs the other way from cost-optimization: intelligence is an unalloyed good, so given the leverage poured behind a decision — capital, code, marketing — you want the best judgment every time, and you stop asking the model you believe is weaker once you can’t tell which answer is right. When a model errs you often don’t notice, which is the argument against trading intelligence for price on consequential work.

The caveat comes from Vercel’s gateway data: frontier intelligence at the right cost and latency wins most production traffic, but cheaper and open models earn their place on verifiable, high-volume tasks — support, browser automation — where a wrong answer is caught cheaply. This is the same fast-default / thinking-when-hard logic applied to model choice rather than reasoning depth: spend intelligence where verification is expensive, economize it where checks are cheap. See The AI Industrial Revolution.

Sources

Open Questions

  • Which tasks in this wiki deserve a thinking model?
  • Should lint passes use a thinking model by default?