Thinking tokens: how reasoning models bill for thought, and when it pays
In brief
Thinking tokens are billed as output, often unseen, and vary by request. How OpenAI, Anthropic and Google bill and cap them, with cost math on current prices.
Steve Elliott · · 7 min read
On this page
Reasoning models bill their thinking as output tokens, at the full output rate, even though you never see most of it. On a request with a short visible answer, thinking was 38% to 96% of the cost in our illustrative example, and the amount changes from one request to the next.
That makes reasoning a pricing decision as much as a model choice. In our view, finance and AI engineering teams should treat the effort setting as a cost control, measure thinking tokens separately in usage data, and pay for deep reasoning only where its error reduction is worth more than the extra tokens. Prices and settings below are as of November 2026.
How are thinking tokens billed?
OpenAI, Anthropic and Google all bill reasoning tokens as output, the pricier side of input and output tokens.
- OpenAI. Its reasoning guide says reasoning tokens are not visible through the API but still occupy the context window and are billed as output tokens.
- Anthropic. Its thinking docs say tokens Claude uses while thinking are "billed as output tokens." Thinking blocks from earlier turns that stay in context are billed again as input.
- Google. The Gemini API pricing page labels each model's rate "Output price (including thinking tokens)". Its thinking guide says pricing is based on the full thought tokens generated.
What you see is not what you pay for. Anthropic returns either a summary of the thinking or nothing, and nothing is the default on its Claude 5 models. It bills the full thinking either way, and its docs warn that the billed output count does not match the visible count. Google returns thought summaries only if you enable them, and bills the full thoughts. OpenAI includes reasoning summaries only if you opt in.
How do providers expose and limit reasoning?
Each provider offers a depth setting and a usage field that counts thinking tokens. The details differ.
| As of November 2026 | OpenAI | Anthropic | Google Gemini |
|---|---|---|---|
| Depth control | Reasoning effort: none, minimal, low, medium, high, xhigh, max, depending on model | Effort: low, medium, high, xhigh, max, depending on model; a fixed thinking budget (minimum 1,024 tokens) only on older models | Thinking level (minimal, low, medium, high, plus xhigh on some Gemini 4 models) on Gemini 3 and later; a token budget on Gemini 2.5 |
| Default | Medium on GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna | High on most models; medium on Claude Opus 5.5 and Haiku 5.5 | Medium on Gemini 3.8 Flash; high on Gemini 3.1 Pro |
| Turning it off | None effort on some models; GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Sol and GPT-6 Astra don't accept it | Not possible on Claude Opus 5.5; possible on Haiku 5.5 at high effort or below | Budget of 0 on Gemini 2.5 Flash; not possible on Gemini 2.5 Pro; minimal reduces thinking but is not described as off |
| Where the count appears | reasoning_tokens in the usage output details | thinking_tokens in the usage output details | thoughtsTokenCount in usage metadata |
The Google entries come from Google Cloud's thinking page, updated November 20, 2026.
Two points matter for cost. First, effort is guidance, not a meter. Anthropic calls effort "a behavioral signal, not a strict token budget," and its older fixed budget was a target rather than a strict cap. OpenAI says its models reason adaptively, using fewer tokens on simple tasks.
Second, the hard caps are per request. OpenAI's max_output_tokens and Anthropic's max_tokens both include reasoning. Anthropic notes that in a tool-use loop each request has its own cap, so the cap does not bound a whole agent turn. OpenAI warns that a request can hit its limit before producing any visible output, so you can pay for input and reasoning and get no answer. It recommends reserving at least 25,000 tokens for reasoning and output when you start.
What does a request with thinking cost?
An illustrative contract clause review: each request sends 6,000 input tokens (the clause and a review playbook) and gets back a 400-token visible finding. The thinking token counts are assumptions chosen to show the range. The model is Claude Opus 5.5 at its list rate as of November 2026: $4 per million input tokens and $20 per million output tokens. Thinking can't be switched off on Opus 5.5, though the model can skip it on simple requests; we assume a clause review always prompts some, so the table starts at 1,000 thinking tokens.
| Thinking tokens | Input cost | Output cost (visible + thinking) | Cost per clause | Thinking share | Per 100,000 clauses |
|---|---|---|---|---|---|
| 1,000 | $0.024 | $0.028 | $0.052 | 38% | $5,200 |
| 4,000 | $0.024 | $0.088 | $0.112 | 71% | $11,200 |
| 16,000 | $0.024 | $0.328 | $0.352 | 91% | $35,200 |
| 40,000 | $0.024 | $0.808 | $0.832 | 96% | $83,200 |
The arithmetic for the 4,000 row: 6,000 × $4 per million is $0.024; 4,400 output tokens × $20 per million is $0.088. The visible finding is the same length in every row. Across the table, the cost per clause moves 16-fold because of tokens nobody reads.
The same request on a smaller model looks different. Claude Haiku 5.5 lists $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens. With 4,000 thinking tokens, a clause costs $0.0006 plus $0.0022: $0.0028, or $280 per 100,000 clauses.
Prices also move. Google's page lists Gemini 3.8 Flash at $0.75 input and $3.75 output through December 31, 2026, and $1.50 and $7.50 from January 1, 2027. The 4,000-thinking request would cost $0.021 today and $0.042 in January on that model, if Gemini thought for the same number of tokens. It might not. Token counts aren't comparable even within one vendor: Anthropic says the tokenizer in Claude 4.7 and later models produces about 30% more tokens for the same text. In our view, per-token prices alone can't compare reasoning models. Measure cost per task on your own traffic; our post on cost per task vs cost per token sets out the method.
Why do reasoning models make AI spend harder to forecast?
Thinking adds variance that a per-token price sheet doesn't show.
- Thinking length varies by request. Anthropic says Claude decides on each request whether to think and how much, and may skip thinking on simple inputs at lower effort. Spend follows the mix of easy and hard inputs.
- The tail drives the average. Illustratively, if 90% of requests think for 2,000 tokens and 10% think for 40,000, the mean is 5,800, and the 10% tail accounts for 69% of thinking tokens. A budget built on the typical request would be about a third of actual thinking spend.
- Defaults change between model versions. Anthropic's docs say a request that omits effort runs one level lower on Claude Opus 5.5 than on Claude Opus 5, and that Sonnet 5.5's levels are calibrated differently from Sonnet 5's. An upgrade can move the bill with no code change, in either direction.
- Earlier reasoning can be billed again. Anthropic keeps prior turns' thinking in context on Claude Opus 4.5 and models numbered 4.6 and higher, billed as input, so long conversations can pay for old reasoning again on each turn, at input or cache-read rates.
- Changing effort can break the cache. Anthropic says a change to the top-level effort level between requests can invalidate prompt caching breakpoints, so a mid-conversation switch to save tokens can raise input cost. Setting effort explicitly to the model's default does not break the cache.
For forecasting, we think the unit to plan on is the distribution of thinking tokens per task, not the mean, refreshed at every model change. The wider method is in forecasting volatile cloud spend.
When is paying for reasoning worth it?
In our view, the test is simple: extra reasoning is worth buying when the errors it prevents cost more than the tokens it uses.
Take the illustrative table above. Moving from 4,000 to 16,000 thinking tokens adds $0.24 a clause. If a missed problem in a clause costs $50 of lawyer time to catch later, the extra thinking pays off only if it cuts the miss rate by at least 0.48 percentage points ($0.24 ÷ $50). At 100,000 clauses a month, the extra thinking costs $24,000; it pays only if it catches at least 480 more problems. Below that, the cheaper setting wins.
Anthropic's own guidance points the same way. It recommends starting Claude Fable 5.2 at high effort, saying most of the capability gain is already there. It says max effort runs slower and costs more than xhigh, often for only a small gain. Research also cautions against assuming more thinking is always better: a paper by Gema and colleagues, presented at ICLR 2026, built tasks where longer reasoning lowered accuracy on 2025-era models, and recommends evaluating models across a range of reasoning lengths.
We think three practices follow:
- Sweep effort on your own evaluations. Run a representative sample at each level and record accuracy, thinking tokens and latency. Anthropic's effort guide gives the same advice for its newest models.
- Route by difficulty. Send routine requests to a small model or low effort, and escalate hard ones. An LLM router or a simple rule on request type can do this.
- Report thinking tokens as their own line. Pull the reasoning count from usage data into cost reports, by application and model, so finance can see when a default change or a harder input mix moves spend.
For how tokens compare with labor and the other costs of an AI initiative, see AI initiative costs: tokens vs labor. In our view, thinking is the part of the token bill most likely to move on its own. As with any unit cost, a thinking-token line only helps if someone owns it.
Thinking-token spend stays explainable when every effort default has a named owner and a review date tied to model upgrades, the kind of decision an operating model should record.
Sources
- OpenAI, Reasoning models, accessed November 22, 2026
- Anthropic, Steering thinking (effort levels, cost control and pricing), accessed November 22, 2026
- Anthropic, Effort, accessed November 22, 2026
- Anthropic, Thinking, accessed November 22, 2026
- Anthropic, Extended thinking (legacy manual thinking budget), accessed November 22, 2026
- Anthropic, Pricing, accessed November 22, 2026
- Google, Gemini thinking (thinking billing), accessed November 22, 2026
- Google, Gemini Developer API pricing, accessed November 22, 2026
- Google Cloud, Thinking (thinking levels, budgets, defaults and usage fields), updated November 20, 2026
- Gema, A. P. et al., Inverse Scaling in Test-Time Compute, arXiv v2, December 15, 2025; ICLR 2026 journal track poster
Frequently asked questions
Are reasoning tokens billed as input or output?
As output, at the model's full output rate, on OpenAI, Anthropic and Google Gemini as of November 2026. Anthropic also bills thinking from earlier turns that stays in context as input: all prior turns on its newer models, only the last turn on others. On the models priced in this post, output costs five times input, so thinking is usually the most expensive part of a reasoning request.
How much does the effort setting change the cost of a request?
There is no fixed ratio. Anthropic calls effort a behavioral signal, not a token budget, and OpenAI says its models reason adaptively. In our illustrative contract review on Claude Opus 5.5, thinking between 1,000 and 40,000 tokens moved the cost from $0.052 to $0.832 a clause. Run each effort level on a sample of your own traffic and record thinking tokens, accuracy and cost.
Is an effort level or thinking budget a hard spending limit?
No. Anthropic calls effort a behavioral signal rather than a strict token budget, and its older fixed thinking budget was a target rather than a cap. The hard limit is the maximum output setting on each request, which includes reasoning. In an agent's tool-use loop, each request has its own cap, so the cap does not bound the whole task.
Should we turn reasoning off to cut costs?
Not by default. In the illustrative example here, going from 4,000 to 16,000 thinking tokens roughly tripled the cost of a clause review on Claude Opus 5.5. That is worth paying only where the errors prevented cost more. Some models, including Claude Opus 5.5, don't allow reasoning to be switched off. Test effort levels on your own evaluations and route easy requests to cheaper settings or smaller models.
Steve Elliott
Published November 22, 2026
More on AI Economics
Tokens, inference cost, cost per task.
Keep reading
Can you capitalize AI development? Internal-use software rules applied to models, prompts and agents
US GAAP has no AI-specific rule, so internal-use software guidance decides. What can be capitalized for agents, fine-tuning, data and tokens, with an example.
Thought Leadership vs People Leadership: Understanding the Key Differences
Thought leaders influence through ideas: original insight that shapes how an industry thinks. People leaders influence through relationships: guiding, developing and motivating a team.
Understanding the Various SaaS Licensing Models
SaaS licensing models include per user, usage-based, tiered, flat-rate and freemium. Learn how each works and how to choose the right one for your product.
5 Examples of Successful Business Models
Five business models that work: subscription, freemium, direct sales, franchise and marketplace or platform. How each makes money, with examples and trade-offs.