Definition
Reasoning effort is Reasoning effort is a setting, often low, medium or high, that tells a reasoning model how much thinking to do before it answers, trading speed and cost for accuracy on harder problems.
A reasoning model works through a problem in a long stretch of hidden or visible thinking before it gives its answer. Reasoning effort is the control that sets how long that thinking is allowed to be. Providers usually offer a few named levels, such as low, medium and high, and sometimes a minimal or maximum level at either end. Some let you set a budget of thinking tokens instead.
Think of the difference between a quick estimate and a worked solution on paper. Low effort is the estimate. High effort is showing all the steps.
The extra thinking is made of tokens, just like the answer itself. A higher setting lets the model write more of them, trying approaches, checking its own work and backing up when it spots a mistake. This is a form of test-time compute: spending more computing while answering rather than during training.
Those thinking tokens usually count toward your bill and take time to produce. So raising the effort generally makes a reply slower and more expensive. The model may also use more of its output limit, so very high effort on a hard task can run into the cap.
Low effort suits simple jobs such as classification, rewriting, short lookups and chat, where speed and price matter and the answer is easy. High effort pays off on multi-step problems: math, hard coding tasks, planning, and agents that must choose actions carefully.
More effort is not always better. On an easy question it can add cost and delay with no gain, and a model can sometimes overthink itself into a worse answer. It also does not remove hallucination, because a long chain of thought can still reach a wrong conclusion.
Measure it. Pick a set of real examples, run them at two or three effort levels, and compare accuracy, time and cost per finished task. Published benchmark scores are often reported at a specific effort level, so check which one before comparing two models, because a score at maximum effort may not match what you get at the default.
Usually yes. The extra thinking is made of tokens that are typically billed and that add to response time.
Use low for simple tasks and high for hard multi-step problems, then test on your own examples to compare accuracy, speed and cost.
Vals AI built 50 web apps with GPT 6 Sol and Claude Opus 5.5 and found agent teams cost 1.8 to 5.1 times as much as one agent, with one significant gain in four setups. Try higher effort on a single agent before you pay for a team.
Meta released Muse Spark Open, an 816B-parameter open-weight model it says runs on two M5 Ultra Mac Studios. The size is real. The scorecard and the full license terms are still missing.
Replit CEO Amjad Masad says OpenAI is now the preferred partner for Replit Agent, citing cost, partnership and performance, though Claude stays in the model selector. Most users never change the default, so the default is the decision.
OpenAI posted a paper on Oct. 26 saying its model refuted the Exponential Time Hypothesis, and Anthropic posted a Claude-found refutation hours later, both checked in Lean. Take the claim seriously and the speedup lightly.
OpenAI released GPT-6.1 Terra on Oct. 26 at $1 per million input tokens and $5 per million output, against $2 and $12 for GPT-5.6 Terra. The middle of OpenAI's lineup just became cheap enough to be your default.