
Do you pay for AI reasoning tokens you never see?
Unity Horizon6 min read
Yes, in full. Reasoning bills at the output rate on all three major APIs, and Claude 5 models return no thinking text by default. Checked 10 November 2026.
Yes, and in full. Thinking tokens bill at the output rate on the Claude API, OpenAI's API and the Gemini API, and none of the three hands back the reasoning it charged for. On Claude 5 models the thinking text is hidden unless you ask for it, so the only trace of the most expensive part of a request is one integer in the usage object.
The meter is split. One request produces two kinds of output, both are priced the same, and you receive one of them.
Where do reasoning tokens land on the bill?
On the output side, which is the expensive side. Claude Opus 5.5 is $4 per million input tokens and $20 per million output, checked by Unity Horizon against Anthropic's pricing page on 10 November 2026. Reasoning is generated rather than sent, so it is billed at the $20 rate, five times what the same text costs to put into the prompt.
The documentation says the quiet part out loud. Anthropic's steering page carries it in a warning box: "You are billed for the full thinking process, not the thinking content visible in the response." Google's thinking page states that response pricing is the sum of output tokens and thinking tokens, and prices the full thought tokens rather than the summary it returns. OpenAI's reasoning guide says reasoning tokens "are not visible via the API" while they "still occupy space in the model's context window and are billed as" output tokens.
The input side has its own counting trap, where the same sentence counts as a different number of tokens at every vendor. That one is the smaller of the two problems, because at least you can count the input yourself before you send it.
Can you read the reasoning you paid for?
Not the reasoning itself, on any setting. What a Claude thinking block contains is a summary, and the documentation shuts the other door in nine words: "No display setting returns the raw chain of thought."
Whether you get even the summary is a per-model default, and the default moved. display: "summarized" was the default on Claude Opus 4.6 and earlier. On Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 and the Fable and Mythos 5 models, the default is display: "omitted": the thinking block comes back with an empty thinking field and an encrypted signature that exists so the next turn can validate it. Switching the summary on changes nothing about the price. The docs put summary generation at no charge, because the charge landed upstream of it.
Picture the engineer who wired a reasoning panel into the internal dashboard on Sonnet 4.6, swapped the model string for Sonnet 5.5 on a Thursday, and is now looking at a column of empty strings with a bill that did not move. The tokens did not stop being generated. The text stopped arriving.
Which field tells you how much of a request was thinking?
One per provider, and each one is a read-only breakdown rather than the number you are charged on. That number is still the single output count, reasoning included.
| Provider | Field to log | What it reports |
|---|---|---|
| Anthropic | usage.output_tokens_details.thinking_tokens |
Raw reasoning tokens generated, always at or below output_tokens. On a streamed response it arrives only on the final message_delta event |
| OpenAI | reasoning count inside output_tokens_details |
Reasoning tokens inside the response's output total |
total_thought_tokens under usage |
Thought tokens the model generated for the response |
Anthropic's own worked example in the docs shows the shape of the problem: 25 input tokens, 348 output tokens, 312 of them thinking. Nine tenths of that response's output charge was reasoning, and under the current default none of it would have come back as text.
Subtract the thinking count from the output count and you have the part of the bill that produced something a person can read. Log that ratio on every request. A monthly average hides the tail, and the tail is what breaks a fixed quote.
Can you still set a thinking budget?
Not on current Claude models. thinking: {type: "enabled", budget_tokens: N} is deprecated, and the extended thinking page states that Claude 4.7 and later models "reject requests that use it, returning a 400 error". Reasoning depth now sits at output_config.effort, which the docs call soft guidance that "doesn't guarantee a token count", with max described as having "no constraint on thinking length". Gemini models think dynamically by default. A search for how to cap reasoning spend still returns guides from this year teaching budget_tokens as the answer, which is a 400 waiting to happen.
What remains is a ceiling rather than an allowance: max_tokens on the Claude API, max_output_tokens on the other two, shared between the reasoning and the answer. Google calls it "a hard cutoff enforced by the infrastructure" and counts thought tokens against it. OpenAI recommends reserving at least 25,000 tokens for reasoning plus output and warns that truncation "might occur before any visible output tokens are produced". That sentence describes a request you pay for and get nothing from. For what the caps above this level can and cannot stop, we costed those separately in Can you put a hard spending cap on an AI agent?.
What does the worst case cost on one request?
Arithmetic over the list prices checked on 10 November 2026, not a measurement: take the 64,000-token ceiling Anthropic recommends at xhigh or max effort and assume one request spends all of it before the answer starts.
| Model | Output price | 64,000 output tokens | 25,000 reserved tokens |
|---|---|---|---|
| Claude Opus 5.5 | $20 / MTok | $1.28 | $0.50 |
| Claude Sonnet 5.5 | $10 / MTok | $0.64 | $0.25 |
| Claude Haiku 5.5 (prompts up to 100k) | $0.50 / MTok | $0.032 | $0.0125 |
A dollar and a quarter is nothing on one request. Ten thousand requests that all hit the ceiling would be $12,800, and no real workload hits it every time, which is exactly why the number nobody has is the distribution. Then it matters who absorbs the variance. A team charging per seat absorbs it quietly. A team that quoted a fixed price per processed document owns it, and owns it on the documents that reason hardest, which are the awkward ones it was already losing money on. We built OFFSET, which files declarations directly to the ΑΑΔΕ portal, and work shaped like that gets priced per declaration rather than per hour.
The vendor made the depth adaptive. The quote did not.
What changes in the code this week?
Log the thinking count next to the output count, per request, and alert on the ratio rather than the total. Then go and look at max_tokens. A ceiling sized for a response with no reasoning in it is the usual reason a reply comes back truncated with stop_reason: "max_tokens", and the truncated half is the half the user wanted.
Pick one effort level for the life of a cached conversation and leave it alone. The resolved effort value is rendered into the prompt, so changing it mid-conversation throws away your cache breakpoints: the script in Anthropic's docs shows a third request re-paying 3,546 tokens of cache creation after a single step down from medium to low.
Last, delete any code that assumes a thinking block will be there. The model decides per request whether to think, at every effort level, and the documentation is blunt about what that means for your parser: "Don't build application logic that assumes every assistant turn starts with one."
We have not published a distribution of thinking tokens per request from a production workload of our own, and the figures above are arithmetic over published prices rather than an invoice. The ratio is the thing worth measuring, and it will differ by prompt shape more than by model.
What share of your last thousand requests was reasoning nobody read, and which field in your logs would tell you? If the answer is that no field would, that is the kind of seam we work on.
Common questions
Are reasoning tokens billed even though you cannot see them?
Yes, at the output rate. Anthropic's documentation states that you are billed for the full thinking process rather than the thinking content visible in the response. OpenAI's reasoning guide says reasoning tokens are not visible via the API while they still occupy context window space and are billed as output tokens. Google prices a response as the sum of output tokens and thinking tokens.
How do you find out how much of an API bill was reasoning?
Each provider exposes one field. On the Claude API it is usage.output_tokens_details.thinking_tokens, which is always less than or equal to output_tokens and appears only on the final message_delta event when streaming. On OpenAI it is the reasoning count inside output_tokens_details. On Gemini it is total_thought_tokens under usage.
Can you still set a thinking token budget on Claude?
Not on current models. The budget_tokens parameter is deprecated and Claude 4.7 and later models reject requests that use it with a 400 error. Reasoning depth is set with output_config.effort instead, which the documentation describes as soft guidance that does not guarantee a token count.
What is the hard limit on what one request can spend on reasoning?
max_tokens on the Claude API, and max_output_tokens on OpenAI and Gemini. Thinking shares that ceiling with the answer text, so a single request can spend the whole allowance on reasoning and return a truncated reply. OpenAI documents that truncation can occur before any visible output tokens are produced.


