Effort
A dial for how much reasoning the model does before it answers. More effort means a better shot at hard problems.
Effort is a dial for how much reasoning a model does before it answers. It is set per request and controls the length of the thinking the model works through before it starts writing the response you see. That thinking is generated at inference time like everything else, the harness often hides it, but it is real work.
Higher effort costs more and runs slower. The reasoning is emitted as tokens, billed as output tokens even when you never see them, and produced one at a time. Turning effort up lengthens the wait before the answer arrives and adds to the bill.
Most harnesses expose effort as a small ladder: low for mechanical edits and well-specified changes, medium as the everyday default, high for tricky bugs and design decisions, max for the hardest problems where a wrong answer is expensive to unwind. Getting it wrong cuts both ways. Too low on a hard problem gives a confident, shallow answer that reads fine and is wrong in a way that costs you later. Too high for a one-line rename means sitting through a long think that produces nothing the lowest setting would not have. Match effort to the task, not to the session.
Related terms
Output Tokens
Tokens the model generates back. Billed at a higher rate, because generating costs more than reading.
Latency
How long you wait. The time to the first token matters more than the total.
Cost per Token
What you pay for each unit of text the model reads or writes.
Inference
Running a trained model to produce an answer. What happens on every request.
Next-token Prediction
What the model actually does. Samples one next token from the context, appends it, and runs again.
Want this working inside your team?
It starts with one conversation about your workflow.