I couldn’t find anything in the documentation about the chat effort levels. Let’s use GPT5.5 chat as an example. We have the following effort levels: None, low, medium, high, very high.
First, I would like examples of when to select each one, as there are 5 levels and I usually think that if the task is simple, I will use low effort and if it is complex, I will use higher effort. But I’m a little lost regarding the 5 levels.
Another question would be about the costs for use. The documentation only provides one cost information: does this mean that regardless of the effort level, the same amount will be charged? If not, what would the costs be for each level?
Very high: big messy tasks, deep investigation, or cases where you want the model to slow down and really work through it
For cost: the published model price is still the per-token price. Effort level does not usually mean a different listed price per token, but higher effort can use more reasoning/output tokens, so the total request can cost more. So “very high” may consume usage faster than “low” even though the pricing table only shows one rate for the model.
Think of reasoning effort as “how much time/budget should the model spend checking its answer before it responds.”
If you choose None, it’s closer to giving the immediate answer. If you choose Very High, it has more room to work through the problem, sanity-check itself, consider edge cases, and make sure the response is actually useful.
A simplified way to picture high amounts of reasoning is:
What’s the answer?
Am I confident that’s the answer?
What edge cases might I be missing?
Am I explaining this in a helpful way?
Is there anything else the user likely needs to know?
That’s not literally the exact internal process, but it’s a pretty good mental model for how the setting affects the response.
So the additional cost is because of the additional tokens it used in considering these options, rather than a higher price simply for selecting the Reasoning Option.