Reasoning models: what thinking tokens are and what they cost

A reasoning model is a language model that works through a problem in tokens before it writes the answer you see. OpenAI's guide says reasoning models "use internal reasoning tokens before producing a response", and Anthropic describes Claude restating the question, trying approaches, checking intermediate results and abandoning paths "that do not hold up". The short version on October 23, 2026: this extra work helps on math, code, analysis and multi-step tasks, and you pay for it as output tokens even when the reasoning is hidden from you. This guide explains how it works in the OpenAI, Anthropic, Google and DeepSeek APIs, what it costs and how to control it, based on each vendor's documentation read on October 23, 2026.

What changes when a model reasons

Anthropic puts the problem well: "A model that answers in a single pass has to get everything right on the first try: no scratch work, no checking, no changing approach halfway through." A reasoning model gets that scratch work. It writes a first plan, tests it, notices a mistake and corrects course, and only then produces the final answer. The result is better on tasks where the answer depends on intermediate steps, which Anthropic lists as "math, coding, analysis, and long-running agentic work".

It is not a guarantee of truth. A model that reasons can still reason from a wrong premise, so the advice in our guide to model hallucinations still applies. And reasoning has a price in time and tokens, which is why every vendor now lets you choose how much of it you want.

Reasoning in four APIs

Table: reasoning controls by vendor (official documentation, checked October 23, 2026)

Two details explain most surprises. On Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, thinking "is already on and needs no configuration", and the thinking text is hidden by default (display is "omitted"). And what Claude, OpenAI and Gemini show you is a summary, never the raw chain of thought: Anthropic states that no setting "returns the raw chain of thought".

You pay for thinking you never see

This is the part that changes budgets:

• OpenAI: reasoning tokens are not visible through the API, still take up the context window and are billed as output tokens. The max_output_tokens limit covers reasoning too, and if a response hits it, the status is incomplete, which can happen before any visible output is produced, so you can pay for reasoning and get no answer. OpenAI recommends "reserving at least 25,000 tokens for reasoning and outputs" when you start.
• Anthropic: thinking tokens are "billed as output tokens, even when the thinking text isn't returned to you". With a summary, "You're charged for the full thinking tokens generated by the original request, not the summary tokens", so the billed count does not match what you see. The usage.output_tokens_details.thinking_tokens field tells you how many there were.
• Google: "response pricing is the sum of output tokens and thinking tokens", billed for the full thoughts even when only a summary comes back.

As an illustration, an answer of 500 visible tokens that needed 4,000 tokens of reasoning is billed as 4,500 output tokens, nine times what you read. Our LLM API cost calculator lets you try your own figures, and the guide to what a token is explains the unit.

How to control depth and cost

• Lower the effort first. Anthropic's advice for "Lower cost or latency" is to "lower effort first. It scales the whole response down, thinking included."
• Set a hard ceiling with the token limit. In Anthropic's words, "Effort is soft guidance. max_tokens is a strict limit."
• Leave room in the context window. Reasoning shares the window with your prompt and the answer, so a long document plus deep reasoning can run out of space. Our guide to context windows covers the arithmetic.
• Keep settings stable when you cache. Anthropic notes that changing the thinking configuration between requests invalidates prompt cache breakpoints.
• Older Claude budgets. Manual budget_tokens has a minimum of 1,024 tokens, Anthropic suggests 16,000 or more for complex tasks and batch processing above 32,000. It is deprecated on the 4.6 models and rejected from 4.7 on, where adaptive thinking replaces it.

Reasoning and tools

Reasoning matters most in agents, because the model can think between tool calls. Anthropic calls this interleaved thinking, and adaptive thinking does it automatically. It also means more state to carry: DeepSeek requires the reasoning_content of previous turns to be passed back in requests that use tools, or it returns a 400 error, and Gemini says that in stateless mode you "MUST" resend thought blocks exactly as received. With Claude's older manual mode, tool use only supports tool_choice set to auto or none. If you are building with tools, start with our explainer on function calling.

When a reasoning model is worth it

Use deeper reasoning for multi-step math, debugging, planning, analysis with many conditions and agent tasks. Use low effort, or thinking off where the model allows it, for classification, extraction, short rewrites and ordinary chat, where the extra output tokens and latency rarely change the result. Anthropic warns that higher budgets bring "diminishing returns that depend on the task", so the only reliable answer is to run your own prompts at two effort levels and compare quality, time and the billed output tokens.