When thinking longer pays off.

A reasoning model is an AI model that works through a problem before it answers. In a hidden scratchpad, it breaks the question down, tries approaches, checks its steps and drops dead ends. That makes it much better at math, planning, debugging and long multi-step tasks.

The catch is that thinking costs time and money. For a quick question or a simple rewrite, it adds delay without a better answer. The real skill is matching the amount of thinking to the problem.

01What is a reasoning model?

Think of a strong chess player. In a familiar position, they move almost instantly. In a sharp one, they stop and calculate several moves ahead, because instinct alone would blunder. Reasoning models give AI that second gear.

Without it, a model answers in a single pass. Anthropic’s documentation puts the limitation plainly: a single-pass model has to get everything right on the first try, with no scratch work, no checking and no changing course halfway through. With thinking, it can restate the question, test an idea, notice a mistake and start again before you see anything. Our plain-English tour of how LLMs work covers the rest of that process.

02How thinking works, without the math

The thinking is ordinary text the model writes for itself, made of tokens like any other output. Tokens are small chunks of text, about three quarters of an English word each.

Three practical consequences follow:

• You pay for it. Anthropic, OpenAI and Google all bill thinking tokens at the output rate, and you pay for the full thinking, not just the part you are shown.

• You mostly do not see it. OpenAI and Anthropic return a summary of the reasoning at most, never the raw chain of thought. Google offers thought summaries too.

• It takes up room. Thinking counts toward the model’s context window, the limited amount of text it can hold at once.

Newer models also decide for themselves how much to think. Anthropic calls this adaptive thinking, and Google’s Gemini adjusts its thinking to the request by default. A simple question may get no thinking at all, while a hard one gets a long draft.

03When thinking pays off, and when it wastes money

The rule of thumb: thinking helps when checking your work is cheaper than being wrong.

The labs’ own guidance agrees. Google’s Gemini docs suggest minimal or low thinking for simple tasks like fact retrieval, and the most thinking for complex coding and math. Anthropic’s effort guide, in its notes for Claude Opus 4.7, warns that the top setting adds significant cost for relatively small gains on most workloads, and can cause overthinking on some less demanding tasks.

Research points the same way. A 2025 study by Apple researchers tested reasoning models on puzzles of rising difficulty. Standard models did better on the easiest puzzles, reasoning models pulled ahead on medium ones, and both collapsed on the hardest.

Headline scores for reasoning models usually come from hard math and science tests. Our guide to AI benchmarks explains what those scores do and do not tell you about everyday work.

04Effort settings: how much should a model think?

Most current models have a dial for how hard they work. It changes how much the model thinks, and on some models how thorough its tool use and answers are. Each lab names and scales it differently.

Anthropic’s docs describe effort as a behavioral signal, not a strict budget. At low effort a model still thinks on hard problems, just less. So a lower setting is usually a safe way to save money on routine work.

1. 01

   Start at the default

   Vendors pick defaults that suit most tasks. Use them until you have a reason to change.

2. 02

   Lower it for simple or high-volume work

   Quick answers, rewrites and bulk processing rarely need deep thinking. A lower level cuts waiting time and cost.

3. 03

   Raise it only when answers fall short

   If the model misses steps on a hard problem, try one level up before you switch to a bigger model.

4. 04

   Check the prompt before the dial

   OpenAI’s guide calls effort a tuning knob, not the main way to recover quality. Vague instructions and missing context cause more bad answers than too little thinking.

Choosing between a bigger model and more thinking is its own decision. Our guide on how to choose an AI model walks through it.

05How to prompt a reasoning model

You no longer need to tell a reasoning model to think step by step, because it already does. OpenAI’s guidance is to give a clear goal, strong constraints and an explicit description of the output, without scripting every intermediate step. Spend your words on what a good answer looks like.

Goal: find why checkout fails for some users with a saved card, and propose the smallest safe fix.
Context: the error log and the payment module are attached. The failures started after Tuesday's release.
Constraints: do not change the database schema, and keep the public API the same.
Done means: the root cause in two sentences, the fix as a diff, and a test that fails before the fix and passes after.

Compare that with “fix the checkout bug”. The short version leaves the model to guess the goal, the limits and when it is finished, and thinking harder does not fix a guess.

06How Cyborb’s Air, Liquid and Plasma map to this choice

Cyborb, our desktop AI agent, offers the same choice as three tiers of its built-in agent. Air is instant and free, “the model you leave running”. Liquid is the balanced everyday worker. Plasma is the deepest reasoning, for the hardest problems.

A sensible way to use them follows the same rule as above. Keep Air for quick questions and small edits, use Liquid for most building and fixing, and switch to Plasma when a bug or plan has resisted a lighter pass. In general, deeper thinking means more tokens and more waiting, so it pays to save it for the hard parts. For more on spending wisely, see how to cut your AI bill.

07FAQ

Next, see how to choose an AI model for the job, or learn what tokens and context windows are.