ClaudeClaude output tokens in the MacBook notch
Output tokens are what Claude generated, and thinking tokens are counted here too.
Thinking tokens are output tokens
This is the part that catches people. When a model reasons before answering, those reasoning tokens are output tokens and they are billed as output tokens, whether or not your application displays them.
A product that enabled extended thinking and saw its output token count jump has not got a bug. It is seeing the cost of the reasoning it asked for.
That makes this figure the one to watch when you turn thinking on or change a thinking budget, because the change lands here rather than in input tokens.
Output is the expensive half
Output tokens are priced well above input tokens on every Claude model. The ratio between this figure and your input tokens is therefore a better guide to your bill than either number alone.
A retrieval-heavy product sends enormous context and generates short answers, so input dominates the count and output can still dominate the cost. Check API spend rather than assuming.
Controlling it
max_tokens caps it, and most applications set it far above what they
render.
The thinking budget is the other lever, and it is the one with real leverage once extended thinking is on. A budget set for the hardest request your product makes is applied to the easiest one too unless you vary it.
What is not here
Tokens spent by web search or code execution tools appear in API spend rather than in this count, because they are billed as tools rather than as model output.
What Claude output tokens means, exactly
Output tokens generated on Messages requests in the window. Thinking tokens are output tokens and are counted here.
A whole number.
How the widget behaves
- Buckets
- hourly, daily and weekly
- Shape
- A figure with a sparkline
- Against last period
- Yes
- Fetched from
- api.anthropic.com
Output tokens from other sources
Notch Metrics reports output tokens from 1 other source. Each defines it its own way, so the widgets stay separate rather than being added together.
Questions
- How is Claude output tokens calculated?
- Output tokens generated on Messages requests in the window. Thinking tokens are output tokens and are counted here. Notch Metrics reports that number, unchanged, in the MacBook notch.
- Can I see Claude output tokens over time?
- Yes. Claude output tokens is bucketed hourly, daily and weekly, so the widget carries a sparkline alongside the headline figure, and a comparison against the previous window.
- What unit is Claude output tokens in?
- A whole number.
- Which other sources report output tokens?
- Besides Claude, Notch Metrics reports output tokens from OpenAI. Each provider defines the number its own way, so the widgets stay separate rather than being added together.
- What does it cost to put Claude output tokens in the notch?
- $9 a month, or $49 a year, with every connector included and no per-integration charge. The seven-day trial needs no card, so you can see your own output tokens before deciding.