Predict the output rate token count when reasoning tokens stay hidden
A single reasoning API call uses 800 input tokens, generates 3,200 hidden reasoning tokens, and returns a 200 token visible answer. The provider never shows you the raw reasoning, but bills reasoning tokens at the output token rate. How many tokens does this call bill at the output rate?
Hidden reasoning tokens bill at the output rate, so this call costs 3,200 + 200 = 3,400 output-rate tokens plus the 800 input tokens charged separately.
Imagine paying a consultant who scribbles four pages of notes and then hands you a one-paragraph answer. The four pages went straight into the shredder, you never read them, but you still pay for every minute spent writing them. Reasoning models work the same way. The model thinks privately, the provider charges you for the thinking, and you only see the final answer. Add the invisible thinking tokens to the visible answer tokens and that total is what shows up on your bill at the higher output rate. Input tokens, the prompt you sent in, are charged at the cheaper input rate on a separate line.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Reasoning tokens are the cost surprise of the 2025-2026 era. A team migrates a workflow from a fast model to a reasoning model, ships, and the first end of month invoice is five to ten times higher than the load-test projection. The cause is almost always the same: the team estimated cost from visible reply length, and the model is now generating thousands of hidden tokens per request that nobody saw on screen.
This deep dive walks through the token accounting that every reasoning-API call uses, the arithmetic on the worked example, and the operational practices that prevent the silent cost bloat failure mode. The numeric answer is small. The mental model that justifies it is what stops a $200 daily test from becoming a $40,000 monthly bill.
Three buckets, two rates
A reasoning-model response carries three distinct token counts in its usage metadata.
- Input tokens. Everything the caller sent: system prompt, tool definitions, conversation history, current user message. Billed at the input rate, which is typically four to ten times cheaper than output.
- Reasoning tokens. The hidden thinking trace the model produced before committing to a visible reply. The provider does not return the raw text in most cases (OpenAI o-series, Google thinking models). Anthropic returns the thinking block but the tokens still bill at the output rate.
- Visible completion tokens. The reply the user actually reads.
The pricing rule that matters: input bills at the input rate; reasoning and visible completion both bill at the output rate. There is no third tier. The provider's invoice line for output covers both the trace you saw and the trace you did not.
This is the single fact the worked example is testing. Once it is internalised, the arithmetic is one addition.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI o-series responses include `completion_tokens_details.reasoning_tokens` for exactly this accounting, separate from `completion_tokens`
- Anthropic Claude Opus 4.7 extended thinking returns thinking blocks whose token count is billed at the standard output rate
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you alert on a silent shift from low to high reasoning effort in production?
Log reasoning_tokens per request, compute a rolling p99, and alert when p99 jumps more than say 50 percent week over week. Combine with a per-route cost dashboard so the alert localises to the workload that drifted.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Forgetting that hidden reasoning tokens count as output and adding the 800 input tokens into the output bucket, both of which break the cost model.
60 second bullets to scan on the way to the call.
Which token buckets a reasoning API exposes in its usage metadata
Which buckets bill at the input rate versus the output rate
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.