OpenAIOpenAI input tokens in the MacBook notch
Input tokens are what you sent, with cached reads and cache writes included, and audio and image inputs counted in their own units.
Cached reads are counted here
The figure includes uncached input, cached reads and cache writes together, which is what the Usage page totals. It is deliberately not what you were billed, because cached input is priced well below fresh input.
That distinction matters the moment you adopt caching. Your input token count barely moves and your input bill falls, which looks like a reporting error until you know the number is measuring volume rather than cost.
For cost, read API spend.
Audio and image inputs have their own units
A model accepting images or audio does not count those in text tokens. OpenAI reports them in their own token units, and this figure passes that through rather than attempting a conversion.
So a multimodal product sees a number here that is not comparable with a text-only one, and should not be divided by request count to get an average prompt size in words.
Input is usually the one you can shrink
Most applications send far more context than they need, because context is easy to add and nobody ever removes it. Retrieval that returns ten documents when three would do, a system prompt that accumulated instructions over a year, a full conversation history replayed on every turn.
Three things reduce it in order of leverage: cache the stable prefix, trim retrieval to what actually gets used, and summarise conversation history rather than replaying it.
Reading it with output
Put output tokens beside it. Input dominating means you are paying to send context; output dominating means you are paying to generate. The fixes are not the same.
What OpenAI input tokens means, exactly
Input tokens on completions requests, cached reads and cache writes included. Audio and image inputs are counted in their own token units, as OpenAI reports them.
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.openai.com
Input tokens from other sources
Notch Metrics reports input tokens from 1 other source. Each defines it its own way, so the widgets stay separate rather than being added together.
Questions
- How is OpenAI input tokens calculated?
- Input tokens on completions requests, cached reads and cache writes included. Audio and image inputs are counted in their own token units, as OpenAI reports them. Notch Metrics reports that number, unchanged, in the MacBook notch.
- Can I see OpenAI input tokens over time?
- Yes. OpenAI input 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 OpenAI input tokens in?
- A whole number.
- Which other sources report input tokens?
- Besides OpenAI, Notch Metrics reports input tokens from Claude. Each provider defines the number its own way, so the widgets stay separate rather than being added together.
- What does it cost to put OpenAI input 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 input tokens before deciding.