What Are Reasoning Tokens? Cost Math for 2026

Short answer: Reasoning tokens are the model's internal thinking, generated before it writes the answer you see, and every major provider bills them at the output rate. A 10,000-token thinking burst costs $0.50 on GPT-6 Astra, $0.20 on GPT-5.6 Sol, $0.25 on Claude Opus 5, and just under 4 cents on Gemini 3.8 Flash. They occupy the context window even though they rarely appear in the response, and you can dial them down: reasoning effort on OpenAI and OpenRouter, a thinking budget on Anthropic, thinking_level on Gemini.

Here is a bill that surprises people. You ask one hard question, the answer comes back two paragraphs long, and the usage line says twelve thousand output tokens. Two thousand of those wrote the answer. The other ten thousand were the model thinking, and on GPT-6 Astra they cost five times what the visible text did.

Reasoning tokens are not a billing error. They are the product of the thing you paid for: models that work through a problem before answering. The catch is that the thinking is invisible in the chat window while being fully visible on the invoice, and the same 10,000-token thought spans a 13x price range across the models people actually use.

What are reasoning tokens?

When a reasoning model receives a prompt, it can generate an internal chain of thought before producing its final answer. Those internal tokens are the reasoning tokens. Different providers describe them differently, but the billing rule is consistent:

• OpenAI: "While reasoning tokens are not visible via the API, they still occupy space in the model's context window and are billed as output tokens."
• OpenRouter, which routes to most of the same models: "Reasoning tokens are considered output tokens and charged accordingly."
• Anthropic reports them in the usage data. The usage.output_tokens_details.thinking_tokens field "reports how many of the billed output tokens were internal reasoning."
• Google's Gemini pricing tables list the line plainly: "Output price (including thinking tokens)."

Three practical consequences follow from that rule. First, reasoning is priced at the most expensive rate category, the output rate, commonly 5x the input rate. Second, reasoning counts against the context window, so a long thinking burst consumes room that a follow-up message might need. Third, you cannot see it by default, which makes it the one part of a bill that resists the usual eyeball check.

What reasoning tokens cost on each provider

Rates read from provider pricing pages on September 15, 2026. Per million tokens, standard short-context tiers. The fourth column applies the output rate to a 10,000-token thinking burst:

Model	Input	Output	Cost of 10K reasoning tokens
GPT-6 Astra	$10.00	$50.00	$0.50
GPT-5.6 Sol	$4.00	$20.00	$0.20
GPT-5.6 Terra	$2.00	$12.00	$0.12
Claude Opus 5	$5.00	$25.00	$0.25
Claude Sonnet 5	$2.00	$10.00	$0.10
Gemini 3.8 Flash	$0.75	$3.75	$0.0375
Grok 4.6	$2.00	$6.00	$0.06

Two rate notes before you build a budget on this table. Sol's $4/$20 is promotional, available at least through November 21, 2026. The Gemini numbers are promotional through December 31, 2026 and are scheduled to double on January 1, 2027, to $1.50/$7.50.

The gap between the top and bottom of that list is the whole story. The same thinking effort that costs half a dollar on a frontier OpenAI model costs less than four cents on Gemini's Flash tier. Neither is wrong; they buy different levels of capability. The mistake is choosing a model for the visible answer quality and paying for its thinking unexamined.

The worked example, monthly

Take a concrete workload: you run 200 hard prompts a month through a reasoning model, and each one consumes 10,000 reasoning tokens plus 2,000 visible output tokens. That is 2 million reasoning tokens a month, and here is what each option charges for them:

Model	Reasoning cost per month (2M tokens)
GPT-6 Astra	$100.00
GPT-5.6 Sol	$40.00
GPT-5.6 Terra	$24.00
Claude Opus 5	$50.00
Claude Sonnet 5	$20.00
Gemini 3.8 Flash	$7.50
Grok 4.6	$12.00

At 500 prompts a month, the same math runs to $250 on Astra, $125 on Opus 5, $100 on Sol, $60 on Terra, $50 on Sonnet 5, $30 on Grok 4.6, and $18.75 on Gemini 3.8 Flash. None of those numbers appear anywhere unless you compute them. The chat window shows answers, not arithmetic.

A sanity check on the ratio: with 10,000 reasoning tokens against 2,000 visible ones at the same output rate, the thinking costs 5x the answer. Estimate your bill from response length and you undercount by that factor.

Why you cannot just check the response

The usual bill-hygiene move, reading the output and estimating, fails specifically here:

• Some reasoning models do not return their reasoning tokens in the response at all. OpenRouter documents this: "Some reasoning models do not return their reasoning tokens," even though they are charged.
• Others return a summary or nothing visible, and the number hides in usage fields: OpenAI reports reasoning within the output token count; Anthropic breaks it out via usage.output_tokens_details.thinking_tokens; Google reports it separately as total_thought_tokens. The names differ, the lesson does not: check the usage object, not the message.

If you are building on a BYOK workspace, your provider dashboard reports the same usage fields per request, so you can audit a sample of real traffic rather than trusting a monthly average. The habit that works is a weekly glance at reasoning volume per model, the same way you would check cache reads and cache writes after enabling prompt caching.

How to control reasoning spend

Every major provider now exposes a dial, though they name it differently:

• OpenAI: the reasoning.effort parameter guides how much the model thinks, with supported values that can include none, minimal, low, medium, high, xhigh, and max depending on the model. Medium is the documented default for most workloads; low favors speed and lower token usage. You can also cap total generation with max_output_tokens, which limits reasoning and visible tokens together, at the cost of responses that can come back marked incomplete when the cap bites.
• Anthropic: manual extended thinking takes a budget_tokens budget. The docs suggest starting near the 1,024-token minimum for simple tasks and 16,000 or more for complex ones, and using batch processing if budgets run beyond 32,000. A newer adaptive mode manages the budget without a fixed number.
• Google: Gemini models expose a thinking_level parameter (low, medium, high). The docs frame low levels as the way to cut cost and latency for workloads where thinking is not the value.
• OpenRouter: a unified reasoning object covers effort, a max token budget, and an exclude flag that hides reasoning from the response. Excluding it does not make it free; the tokens are still generated and billed.
• xAI: Grok 4.6 lists "configurable reasoning" among its features, consistent with the rest of the field.

Tactics that pay off faster than parameter tuning: match effort to the task instead of setting one global level, keep routine work (summaries, formatting, short rewrites) on non-reasoning or low-effort models, and re-check after model upgrades, since the same prompt can produce a very different amount of thinking on a newer model. Our BYOK cost math bills a per-message reasoning count at the model's output rate alongside cache assumptions, so you can see the thinking share of a bill before it arrives.

When paying for thinking is worth it

Not every prompt needs 10,000 tokens of contemplation, and the honest version of this article says so:

• Thinking is the product on hard tasks. Multi-step debugging, long-document analysis, and planning benefit from it directly; that is what the premium output rates buy.
• Turning the dial too low has its own cost. A model that under-thinks a hard task tends to miss, and a wrong answer you must re-run can cost more than the thinking you skipped. The cheap setting is only cheap when it succeeds.
• There is no universal best setting. The right effort level depends on your task mix, and the only reliable way to find it is running your traffic and reading the usage data. Start at the default, sample a week of real requests, and adjust from evidence.

The bill will always show two kinds of output tokens. The question worth asking monthly is which kind you paid for: the answer, or the thinking. If you want to run reasoning and non-reasoning models from every provider in one workspace, see how it works, and measure your own numbers against the monthly cost math.

FAQ

Are reasoning tokens billed as input or output tokens? Output. Every major provider treats internal reasoning as generated output, which is why it carries the highest per-token rate on the price list. The rule is stated explicitly by OpenAI, Anthropic, Google, and OpenRouter.

Can I turn reasoning off completely? Sometimes, but only on models that support it. OpenAI's effort parameter includes a none option on some models, and OpenRouter's reasoning controls allow disabling where the model permits. On models where reasoning is mandatory, the wallet-friendly move is choosing a different model for that workload rather than fighting the setting.

Do reasoning tokens count toward the context window? Yes. They occupy context just like visible tokens, OpenAI's documentation says so directly, which means a long thinking phase leaves less room for the rest of the conversation and can push you into long-context pricing tiers on models that charge more beyond a threshold.

Why did my bill land higher than my own estimate? Most estimates are built from visible input and output. Reasoning tokens are generated output you never read, so an estimate that skips them undershoots by however much the model thought. Check the usage fields for a week and build your per-prompt assumption from there, the same discipline covered in what the ChatGPT API costs per month.

Does a cheaper model with reasoning cost less than a premium model without it? Often yes, and that is the useful comparison. A Flash-tier model doing its own thinking can still land far under a frontier model's reasoning bill for similar work. The trade is capability on the hardest tasks, not a flat rule, which is the kind of comparison the honest BYOK math is built for.

---

Start your free trial → — 7 days, all providers, no credit card required.