Reasoning Effort
Reasoning Effort is a setting in modern AI systems that determines how long and how thoroughly the program thinks before giving its answer. More effort usually means better results on hard tasks, but costs more time and money.
Some AI programs that write text don’t produce their answer right away. First, they generate a kind of scratch pad for themselves: intermediate steps, calculations, discarded ideas. Only afterward do they formulate the result for the user. Reasoning Effort is the dial that determines how extensive this scratch pad is allowed to become. Translated literally, the term means something like “thinking effort.” Many providers offer it as a simple choice with the levels low, medium, and high.
Why a dial for thinking effort is needed at all
Most requests to an AI system are harmless. “How many inhabitants does Hamburg have?” doesn’t require long deliberation. A math competition problem or a broken piece of code very much does. Without a dial, you would have to settle on one setting and would always fare poorly in one of the two cases.
The reason is that these intermediate steps cost real money. An AI provider bills based on text chunks that the model processes and generates. The notes count toward this, even though the user often never gets to see them. A task run at a high level can therefore easily cost ten times as much as a task run at a low level.
Then there’s the waiting time. At a low level, the system often responds within a few seconds. At a high level, it may compute for a minute or longer. For a chat on a website that’s unusable, for an overnight data analysis it’s perfectly fine.
What the dial actually changes
The dial doesn’t make the model smarter. It remains the same program with the same learned abilities. What changes is only how much room it gives itself for its intermediate steps. You can compare it to an exam: the same student writes a different solution with twenty minutes of working time than with two hours. Her knowledge is identical, her thoroughness is not.
Technically, this happens via a budget of text chunks for internal deliberation. At a high level, the model may, for example, try out several solution paths and pick the best one. At a low level, it usually takes the first path that seems plausible. Instead of the three levels, some providers also allow a direct number, such as a maximum of 8,000 chunks for thinking.
A common misconception is that more effort is always better. That’s not true. On simple questions, a lot of deliberation can lead the model to discard a correct first answer again. It essentially talks itself into a mistake. Experts call this behavior overthinking.
Where the levels show up in products and reports
Anyone addressing an AI through a programming interface will often find a field there called reasoning_effort. Developers set it depending on the task. A search function in an online shop runs at a low level, a tool for checking program code at a high one.
The same principle also exists in ordinary chat apps, just without the English name. Buttons like “think longer,” “Deep Research,” or a dedicated mode for difficult tasks are, at their core, the same thing. It’s usually available in paid subscriptions because it’s more expensive to run.
In news about AI, the term comes up in connection with leaderboards. When a manufacturer advertises top scores in math or programming, the fine print often says “high reasoning effort.” This is important to know: everyday operation doesn’t run with the same setting. Anyone comparing the numbers of two models should therefore always check which level was used for the measurement.