What makes a model a reasoning model
A standard model answers in a single pass: whatever working it needs has to happen implicitly as it writes. A reasoning model is trained to think first. Anthropic describes the behaviour concretely: the model restates what is being asked, tries approaches, checks intermediate results and abandons paths that do not hold up, and only then writes the response. OpenAI describes its reasoning models as breaking down the prompt and considering multiple approaches in internal reasoning tokens.
The thinking is mostly hidden. OpenAI does not expose raw reasoning and offers an optional summary; Anthropic returns a summary or nothing; Google returns optional thought summaries. In every case you are billed for the full thinking, not the summary.
What reasoning costs
Reasoning is billed as output, at the output rate, by OpenAI, Anthropic and Google. OpenAI notes that its models may generate anywhere from a few hundred to tens of thousands of reasoning tokens per response, depending on the problem — so the cost of a request is no longer predictable from the length of the answer you read. It is also slower: the user waits while the thinking is generated.
Many current flagship models reason by default. Anthropic's newest Claude models have thinking on with no configuration, and some of them do not allow it to be switched off entirely; OpenAI lets you turn effort down, but not to zero everywhere: GPT-6 Astra rejects none, and GPT-6.1 Sol rejects both none and minimal. Treat reasoning as the default cost of a modern frontier model, and budget for it.
Worked exampleA coding question with and without deep thinking
5,000 requests a month on GPT-6 Sol, priced at its current list rates, checked on 2026-09-30.
The reasoning volumes are illustrations within the range OpenAI documents. The question and the visible answer are the same size in both columns; the difference is thinking billed at the output rate.
Change the workload or the model in the LLM API cost calculator.
Controlling it
Vendors expose a soft dial and a hard limit. The dial is reasoning effort — OpenAI's reasoning.effort, Anthropic's effort, Google's thinking_level — which shifts how much the model thinks on average. The limit is the output-token cap, which covers reasoning and answer together. Anthropic recommends lowering effort first to cut the cost of a thinking workload, and using max_tokens when you need a strict ceiling.
When a reasoning model is worth it
Use one where an extra minute of thought changes the answer: debugging, refactoring across files, maths, planning a multi-step task, reviewing a contract clause against several rules. Skip it — or run it at low effort — for extraction, classification, short rewrites and chat replies, where it adds cost and latency without a better result. Routing requests by difficulty lets one product use both.