Effort is a setting in some AI models that steers how much work they do before answering: how much they reason and, depending on the provider, also how long the answer is and how many tools they call. More effort usually costs more and takes longer, but it guarantees no token count, no time and no better quality.
A question from the app. Pick an answer.
What does effort steer?
It does not guarantee it: more effort can help, but you have to measure it on your cases.
It affects reasoning and, depending on the model, also the answer or tool use. It steers, it does not count.
It steers, it does not fix: two requests at the same effort can spend differently.
It is a dial that steers, not an exact command. With more effort, a reasoning model usually writes a longer draft before answering and, depending on the provider, it may also lengthen the answer or, in an agent, call its tools more often. With less, the opposite.
All that extra work is tokens being generated. Part of it is reasoning you do not see and still pay for. The price per token does not change, but there are more tokens, so the bill goes up.
Steering is not fixing. Two requests at the same effort can spend different tokens and time, and more effort does not ensure a better answer. Levels also depend on the model: what a provider calls "high" in one model is not the same as "high" in another, and not every model has this dial.
The default level changes too. In the Anthropic API, as of October 2026, Claude Opus 5.5 and Claude Haiku 5.5 start at "medium" and the other models at "high": if you do not choose, switching models can change the bill without you touching anything else. And it is no longer a big model thing: Haiku 5.5, the smallest of its family, has the dial too.
It tends to help with logic, math, hard code and multi-step plans, where each step depends on the previous one and a draft prevents mistakes. On simple tasks, like deciding whether an email is spam, it can cost more without being more right. There is no fixed rule for translating, summarizing or classifying: it depends on your cases.
Since nothing is guaranteed, the way to choose is to test. Take a sample of your real cases, run it at different levels and compare cost, time and hits. Keep the lowest level that is right often enough. "Always the highest, just in case" is paid on every call. To see how the bill adds up, check the other page.