Reasoning token pricing means every major AI provider — OpenAI, Anthropic, Google, and DeepSeek — bills the model’s hidden internal “thinking” before it writes an answer at the same rate as the visible output, even though you never see most of it. A request that returns a 300-token answer can carry 3,000 or more billed reasoning tokens underneath it, and prompt caching doesn’t reduce that cost, since reasoning tokens are freshly generated on every single turn rather than reused. Here’s how to actually find that number for your own workload before it shows up as a surprise on a bill.
The pricing page answers a different question than the one that matters
A provider’s pricing page tells you the dollar rate per million input and output tokens. It doesn’t tell you what fraction of a typical request on a reasoning-enabled model will be invisible reasoning rather than the answer you actually asked for — and that fraction, not the sticker rate, is usually what decides whether a model is affordable for your workload. Scaled up, the difference is not trivial: a workflow running 10,000 requests a month, each carrying roughly 3,000 billed reasoning tokens on top of a 300-token visible answer, bills for about 33 million total output-rate tokens a month — not the 3 million a naive per-response estimate would suggest.
How each major provider actually handles this
| Provider | What happens to reasoning tokens | Worth knowing |
|---|---|---|
| OpenAI | Hidden by default; billed as output under output_tokens_details.reasoning_tokens | A detailed breakdown from OpenAI’s own community forum shows a real worked example: 500 regular input + 2,000 cached input + 3,000 reasoning + 300 visible-answer tokens billed out to roughly 5,800 total tokens for one request |
| Anthropic | Visible as “thinking” blocks, still billed at the output-token rate | Anthropic’s own documentation states plainly that the billed output count won’t match the visible token count you see in the response |
| Google Gemini | “Thinking tokens billed at the same rate as output tokens,” per Google’s own pricing documentation | Billing reflects the full raw thought process the model generated, even when the API surfaces only a short summary or nothing at all of it to you |
| DeepSeek | Thinking mode is on by default and adds extra output tokens | A second, separate gotcha stacks on top: DeepSeek’s rates double during a defined daily peak window, so time-of-day and reasoning volume can compound |
Three cost levers almost nobody checks before they get surprised
- Prompt caching doesn’t touch reasoning tokens. Caching discounts repeated input — a system prompt or shared context you send on every call — but reasoning tokens are freshly generated output on every single turn, so there’s no discount lever available for them at all, regardless of how repetitive your prompts are.
- Reasoning tokens still consume context window space. Even though you never see them, they count against the model’s context limit like any other generated token. A conversation that comfortably fits inside one model’s window can hit the ceiling on a higher-reasoning-effort model purely from invisible thinking, with no change to your actual prompt.
- Promotional pricing has an expiration date you need to track. Google’s own current Gemini 3.8 Flash pricing — $0.75 input, $3.75 output per million tokens — is a promotional rate that runs only through December 31, 2026, after which it doubles to $1.50/$7.50. Building a dependency on today’s rate without checking whether it’s introductory is a common way a cost projection quietly goes stale.
What to actually do about it — five checks, not a spreadsheet
- Find the reasoning-token field in your provider’s usage output —
output_tokens_details.reasoning_tokensfor OpenAI, the thinking-token count in the usage payload for Anthropic and Google — and start logging it alongside your regular input/output counts from your very first production request, not after a bill surprises you. - Run one representative real task and compute your own multiplier. Generic guidance suggests budgeting 3–5x your expected visible output for a reasoning model, but that range varies enormously by task type — measure it on your own workload rather than trusting an industry rule of thumb.
- Test a lower reasoning-effort setting before assuming you need the highest one. Most providers expose a tunable effort parameter (OpenAI’s reasoning.effort, Anthropic’s effort parameter, Google’s own thinking configuration) — the same setting our Qwen vs Claude Opus 5 comparison flagged for quietly defaulting to a high-cost tier, and the same tuning question our coding-agent harness guide recommends testing directly on your own tasks rather than assuming the maximum setting is necessary.
- Check whether the rate you’re budgeting against is promotional before committing to it long-term — the Gemini 3.8 Flash example above, also covered in our Muse Spark vs Gemini comparison, is exactly the kind of rate worth re-checking against the provider’s own current pricing page rather than a comparison article, including this one, that may already be out of date by the time you read it.
- Add cloud-platform markup if you’re not calling the API directly. Routing through Azure, AWS Bedrock, or Google Cloud Vertex AI typically adds a further markup over the vendor’s own direct API price — factor that in before comparing providers head to head on list price alone.
Where this connects to decisions you’ve likely already made
This is the same underlying issue behind a hidden cost called out in our Muse Spark 1.3 review — Meta’s separate contributor, standard, and max pricing tiers are one version of the same “the sticker price isn’t the real price” problem — and in our GPT-6 Astra vs Claude comparison, where a 2.5x list-price jump only tells part of the story without knowing the reasoning-token volume behind it. If you’re deciding whether a given model is actually worth its price for your business, this calculation belongs next to the framework in our guide to free vs. paid AI tools and our guide to cutting AI costs for small teams — both are incomplete without accounting for what a reasoning-heavy workload actually bills, not just what the plan or rate card advertises.
Who doesn’t need to worry about any of this
If you’re paying a flat monthly rate — a ChatGPT Plus, Claude Pro, or Gemini Advanced subscription rather than metered API access — none of this changes your bill. Reasoning-token billing is specifically a pay-per-token, API-level concern; a flat consumer subscription doesn’t itemize or surface it because you’re not being charged per token in the first place.
What are reasoning or thinking tokens, and why am I billed for them?
They’re the model’s hidden internal processing before it writes a visible answer — used to plan and reason through a problem. Every major provider bills them at the same rate as visible output tokens, even though you typically never see the actual reasoning text.
How much extra should I actually budget for a reasoning model?
A reasonable starting estimate is 3–5x your expected visible output, but the real multiplier varies significantly by task. Run a representative request through your own workload and check the reasoning-token count in the usage response rather than relying on a generic rule of thumb.
Can I see how many reasoning tokens my request actually used?
Yes. OpenAI reports this under output_tokens_details.reasoning_tokens in the API usage object; Anthropic and Google include a thinking-token count in their own usage payloads. DeepSeek is currently the only major provider that exposes the raw reasoning text itself rather than just a token count.
Does prompt caching reduce reasoning-token costs?
No. Prompt caching discounts repeated input tokens only. Reasoning tokens are generated fresh as output on every single turn, so there’s no caching discount available for them regardless of how repetitive your prompts are.
Does any of this affect me if I just use a ChatGPT Plus or Claude Pro subscription?
No. This only applies to metered, pay-per-token API or platform billing. A flat-fee consumer subscription doesn’t itemize reasoning tokens separately because you’re not being charged per token to begin with.
Why would my bill go up even though the provider’s per-token price didn’t change?
The most common cause is reasoning-token volume — a more complex task, a higher reasoning-effort setting, or a model update can all increase how much hidden reasoning a request generates, which raises your bill even when the published rate per million tokens stays exactly the same.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
