What are reasoning tokens (in plain English)?
Plain-English definition · Updated 2026-11-01. Billing rules are quoted from OpenAI's and Anthropic's own documentation; the arithmetic on their examples is ours.
AITop is an independent guide. Prices and features change fast — always check the vendor's page.
Reasoning tokens (also called thinking tokens) are the tokens a reasoning model writes for itself — planning, checking, backtracking — before and between the text you actually see. They are billed at the full output-token rate, and they are usually hidden from the response. OpenAI's reasoning guide states it plainly: reasoning tokens "are not visible via the API" but "still occupy space in the model's context window and are billed as output tokens" — anywhere from a few hundred to tens of thousands per response. Anthropic's documentation agrees on the meter ("thinking tokens are billed as output tokens") and its own published usage example bills 348 output tokens of which 312 are thinking — 90% of that output bill bought text nobody read (our arithmetic on their example). This is why a one-paragraph answer from a reasoning model can cost several times what its length suggests, and why the dial that actually controls the bill is the reasoning-effort setting, not your prompt.
What it actually means
A normal model goes prompt in, answer out. A reasoning model inserts a step: it generates a private draft — break the problem down, try an approach, check it, discard it, try again — and only then writes the visible answer. Those draft tokens are reasoning tokens. They consume real computation, they occupy the model's context window alongside your prompt and the answer, and the two largest providers meter them on the same line as visible output. OpenAI's guide is explicit that a response "may generate anywhere from a few hundred to tens of thousands of reasoning tokens" depending on how hard the model decides the problem is — the model chooses, not you.
Where the number lives matters, because it is the difference between an explainable bill and a mystery. OpenAI's Responses API reports usage as output_tokens with an output_tokens_details.reasoning_tokens breakdown, and its documentation example bills output_tokens: 1186 with reasoning_tokens: 1024 — the visible answer is roughly 162 tokens, so 86% of that output bill is invisible text (our arithmetic). Anthropic reports output_tokens with output_tokens_details.thinking_tokens; its documentation example bills output_tokens: 348 with thinking_tokens: 312 — 90% thinking. Two providers, two APIs, the same shape: the authoritative billing number is the total, the breakdown is observability, and the gap between them is the part of the bill most buyers never see.
One structural detail is worth understanding because it explains the pricing: reasoning tokens are output, and output has no discount tier. As our prompt caching entry explains, repeated input prefixes can be billed at a tenth of list — but reasoning spend is paid at full output price, every time, with no read-back mechanism. A model that burns 20,000 reasoning tokens to produce 500 visible ones has written 40× its visible length into the most expensive segment of the meter.
Why it matters when you're picking a tool
The visible answer is no longer a cost estimate. Our token entry explains that output bills at several times the input rate; reasoning tokens make that ratio worse by inflating the output count invisibly. When you compare two chatbots on price per "message", you are often comparing different amounts of hidden thinking. A model that answers a simple question after 200 reasoning tokens and one that burns 3,000 on the same question can look identical in a demo and differ by an order of magnitude on an invoice. If a pricing page or a review does not say whether a model reasons by default, you do not know its real output price.
The control is the effort dial. Both major providers now expose a reasoning-effort setting rather than an on/off switch — OpenAI's guide documents effort levels from low up through xhigh and max (with some current models not supporting none at all), and Anthropic's current models use an equivalent effort ladder. Lower effort favors speed and fewer tokens; the top of the ladder is for work your own evals show benefits. The practical buyer takeaway: a tool that lets you set that dial is cheaper to operate than one that hides it, because the dial — not your prompt — is what moves this bill.
Know the failure mode that costs money for nothing. OpenAI's documentation warns that a response can hit max_output_tokens "before any visible output tokens are produced, meaning you could incur costs for input and reasoning tokens without receiving a visible response" — and recommends reserving at least 25,000 tokens for reasoning and outputs when starting out. Read that in price terms: a mis-set cap can bill a full output load and return nothing. On Anthropic's side, thinking tokens are "a subset of your max_tokens parameter" and count toward rate limits, so an undersized budget produces the same outcome. This is also why our inference cost entry tells you to budget by bill segment, not by "a query".
The 2026 reality check
Reasoning is now the default, not an opt-in. What began as a specialty mode ("extended thinking") has become the standard operating mode of frontier chatbots and coding assistants — several current models think on every request and no longer let you switch it off entirely, only steer it down. That is good for answer quality and quietly bad for bills: the invisible line item exists on every request whether or not the task needed it. Neither provider discounts reasoning tokens below the output rate, and hiding them (OpenAI returns summaries of the chain, Anthropic can omit it entirely on some models) does not make them free — Anthropic's docs state the billing is "the same regardless of the display setting".
The published examples are the honest ones — read them. OpenAI's guide example (1,186 billed, 1,024 reasoning) and Anthropic's steering example (348 billed, 312 thinking) are the vendors' own numbers, not our estimates, and both land in the same place: 86–90% of the output bill was invisible thinking on a short, easy task. Real workloads vary widely — long agentic tasks can amortize reasoning across many visible outputs, and trivial requests sometimes still spend hundreds of tokens deciding they are trivial. The ratio you should trust is the one in your usage logs, which is why every serious AI tool now exposes reasoning_tokens / thinking_tokens in its usage object.
Where it shows up in tool pricing. This entry is one of the three reasons (with cache discounts and batch discounts) that two tools quoting "the same" per-million price can bill completely differently for the same job. Coding assistants and agents are the heaviest reasoning spenders — which is part of what you are actually comparing in our best AI coding assistants ranking: not just the model, but how much thinking the harness lets the model do per action, and whether you can cap it. On consumer chat plans, reasoning mostly hides inside rate limits and "message caps" — the plan that throttles you after 10 messages may simply be the one whose model thinks harder per message.
Quick checklist
- Read the usage object, not the transcript. Log
reasoning_tokens(OpenAI) orthinking_tokens(Anthropic) per request from day one. The visible answer's length tells you nothing about the bill. - Compute your invisible share once. thinking ÷ total output is the number that explains why bills run past estimates. On the vendors' own examples it is 86–90%; your workload will differ — measure it.
- Set the effort dial deliberately. Low effort for classification, extraction and routing; high only where your evals show it pays. On models where reasoning can't be disabled, the dial is the only lever.
- Cap output with reasoning in mind. An undersized
max_output_tokenscan bill a full load of thinking and return no visible answer at all — OpenAI suggests reserving at least 25,000 tokens for reasoning plus output when starting out. - Compare tools on the same reasoning regime. A cheaper per-token price means little if the tool's default effort burns twice the tokens; normalize comparisons to cost per completed task, not per million tokens.
Where you'll hit it
Every bill on this site passes through this line item. Our AI token entry covers the metered unit itself, inference cost explains why output bills at several times input, and prompt caching is the one discount that never applies to reasoning spend. When you are choosing between the best AI chatbots or reading our Perplexity vs ChatGPT comparison, part of what you are comparing is how much invisible thinking each product's models do per answer. Related terms: KV cache is the VRAM the hidden chain occupies while it runs; tokens per second slows down while the model streams its thinking. Full list on the AI tools glossary.
The bottom line
Reasoning tokens are the part of an AI bill you pay for and never see — billed at the full output rate on both major platforms, occupying context, and chosen by the model rather than by you. The vendors' own documentation examples put the invisible share of short answers at 86–90% of the output line. None of this makes reasoning models a bad buy; it makes reasoning-blind buying expensive. Read the usage object, set the effort dial like the budget control it is, cap outputs with the reserve the docs recommend, and normalize every tool comparison to cost per completed task. The thinking was always on your invoice. Now it's a line item you can price.