How the effort setting works
Effort is guidance, not a budget. OpenAI describes reasoning.effort as guiding the model on how much to think: lower effort favours speed and fewer tokens, higher effort lets the model think more completely. Anthropic's effort controls how much work Claude puts into the whole response — in adaptive thinking mode, including how often and how deeply it thinks. Google's thinking_level does the same for Gemini.
The levels differ by vendor and by model. OpenAI lists none, minimal, low, medium, high, xhigh and max as possible values, while noting that support is model-dependent: GPT-6 Astra rejects none, and GPT-6.1 Sol accepts neither none nor minimal and defaults to medium. Anthropic's levels run from low to max, and on some of its newest models certain thinking settings are accepted only at high effort or below.
What each step up costs
Effort changes the number of reasoning tokens, and those are billed as output. Because the count is decided by the model per request, no vendor publishes a fixed token figure per level — OpenAI only describes a range from a few hundred to tens of thousands of reasoning tokens per response. The practical way to price a level is to run a sample of your own requests at each setting and read the output tokens from the usage report.
Changing effort can also cost you cache hits. Anthropic renders the thinking configuration and the effort level into the prompt itself, so switching levels starts a new cache prefix. OpenAI's GPT-6 models offer a configuration update that changes effort mid-conversation while keeping the original prompt prefix intact for caching.
Worked exampleOne request at three effort levels
20,000 requests a month on GPT-6 Sol, priced at its current list rates, checked on 2026-09-30.
The reasoning counts per level are assumptions for illustration — vendors do not publish fixed figures, and the real count varies per request. Measure your own at each level before choosing.
Change the workload or the model in the LLM API cost calculator.
Choosing a level
Start at the vendor's default and move down until quality drops on your own evaluation set; then go back one step. OpenAI suggests trying low before none for latency-sensitive work, and reserving xhigh for deep research and long agentic runs where evaluations show a clear benefit. Anthropic's guidance for a thinking workload that costs too much is to lower effort first.
Effort is soft, so pair it with a hard limit. A sensible output cap stops a rare runaway response; effort shapes the average. Different steps of one product can use different levels — a quick reply at low effort, a planning step at high — which is a form of routing within one model.