Controllable Effort

When you ask a text question to a computer program that formulates answers, that program can work quickly or thoroughly. Quickly means: it starts writing immediately. Thoroughly means: it tries out several solution paths in the background, discards bad ones, and checks the result once more. Controllable Effort means that you can set this difference yourself. You essentially turn a dial between “fast and cheap” and “slow and careful.” The dial doesn’t change the program itself, only how much work it puts into each question.

Why a dial is better than a fixed setting

Not every question deserves the same amount of effort. “How many federal states does Germany have?” can be answered in a tenth of a second. A mathematical proof or a longer programming problem, on the other hand, requires many intermediate steps. A system that always thinks at maximum capacity wastes money and time on simple questions. A system that never thinks fails on the hard cases.

For companies, this is primarily a matter of cost. Every answer from an AI costs computing time in a data center, and that computing time is paid for. If a company handles millions of requests daily, it makes a huge difference whether an average answer requires 200 or 5000 computation steps. With an effort dial, the expensive variant can be reserved for the few difficult cases.

Then there’s waiting time. Users abandon a chat if it stays silent for twenty seconds. For a search query in a browser, speed matters more than ultimate precision. For a medical or legal analysis, it’s exactly the opposite. The dial allows the same model to be used for both situations.

What the model does with the additional effort

Modern language models often generate an internal text before the actual answer, in which they break down the problem step by step. This intermediate text is called a chain of thought. The more effort that is allowed, the longer this chain may become. At a low level, it is heavily shortened or skipped entirely.

Technically, the dial is usually implemented as an upper limit for text units. A text unit, in technical jargon a token, is a word fragment of a few characters. A setting like “low,” “medium,” or “high” determines how many such units the model is allowed to use for its internal thinking. Some providers instead allow a specific number to be set, such as a maximum of 8000 units.

A second approach is to have the same question answered multiple times and compare the results. Five runs cost roughly five times as much, but deliver a more stable answer. It’s important to address a common misunderstanding: more effort doesn’t make the model smarter, only more careful. Knowledge that was never learned during training does not emerge through longer deliberation either. And beyond a certain point, additional effort barely improves results any further.

Effort levels in today’s products

In chat interfaces, the dial appears as a selection menu. There you’ll find labels like “fast,” “think,” or “research deeply.” If you choose the thorough level, you often see a note that the answer will take longer. With some providers, the expensive level is additionally limited per day.

For developers, the setting is a parameter in the programming interface. A team can use it in the code to specify that standard requests run cheaply and only flagged special cases get the high level. Some systems now decide this automatically: a small preliminary model estimates the difficulty of a question and selects the appropriate level itself.

In news about AI, you’ll usually encounter this term in connection with costs and benchmark tests. Providers often report best scores for the highest effort level, even though a lower one runs in everyday use. Anyone reading numbers from such tests should therefore always check which level was used for the measurement.