The Usage & cost view
Read company-wide model usage and spend by Coworker, Knowledge Base, system job, team and person, for any month or custom date range.
Usage & cost shows what your Coworkers, Knowledge Bases and system jobs have used, taken
from the record of every model call. It is a view on the Governance page, selected with the Usage & cost button above the period
control (the other view is Reviews). Every signed-in
member can open it. People with the people_manager or executive (Executive
(view-only)) ability see usage across all Teams, so the numbers cover the whole company. People
with manage_coworkers or manage_apps see the usage of the Coworkers they supervise or the
Apps they built. Everyone else sees only the usage of the teams they belong to, across all four
tables and the totals. Usage with no team is hidden from them, and a person who is on no team
sees no usage.
The record of a model call holds counts, cost and timing only (who made the call, which model, how many tokens, what it cost). It does not keep a copy of the prompt or the reply: a chat call points at the messages already stored in that conversation, and calls made with an API key or by a background job carry no content at all.
Choose a period#
Under the Usage & cost and Reviews buttons, a period control offers three choices:
- This month: from the first of the current month to the first of next month.
- Last month: the previous calendar month.
- Custom: two date fields, From and To. The To date is included in full.
Months are calculated in your browser's time zone. In Custom, the page asks you to "Choose a start and end date (the end on or after the start)." until both dates are valid.
Totals#
Above the tables, five tiles summarize the period:
| Tile | Meaning |
|---|---|
| Total cost | Spend in US dollars. Amounts under one cent read <$0.01. See How cost is calculated. |
| Calls | Number of model calls. |
| Input tokens | Tokens sent to models. |
| Output tokens | Tokens returned by models. |
| Errors | Calls that failed. |
Unpriced calls#
Some models have no known price (see How cost is calculated). The view never shows those as $0.00:
- If every call in the period is unpriced, Total cost reads Unpriced and shows how many calls that is.
- If some are, Total cost shows the amount for the calls with a known price and notes "Excludes N calls with no price".
- In each table, a row whose calls are all unpriced shows an Unpriced label in the Cost column. A row with only some shows its amount followed by "· N unpriced".
Token and call counts include unpriced calls, so you can still compare volume. An admin can give an unpriced model a price under Settings then Model prices; see Price models the price list cannot price.
If nothing was recorded, the page shows "No usage in this period". The tiles and tables appear as calls happen. If the page cannot load, it says "Usage couldn't be loaded. Try again in a moment."
Breakdown tables#
Each table has the columns Calls, Input tokens, Output tokens, Errors and Cost, sorted by cost with the highest first. Rows that are entirely unpriced have a cost of zero in this ordering, so they sort last. A table with no rows says "No usage in this period."
| Table | One row per | Notes |
|---|---|---|
| Coworkers | Coworker | Chat and conversation-compaction calls made for that Coworker. |
| Knowledge Bases | Knowledge Base | Ingestion and search calls. A row named Search embedding collects search queries, and a row named Search rerank collects the relevance-ranking calls made for searches (the default small model, or the company default model); neither is billed to one Knowledge Base. Reading scanned pages and pictures is billed to the Knowledge Base the file was added to, and a row named Image-reading test collects the Test image reading calls. No Knowledge Base recorded collects ingestion and image-reading calls that were not tied to one. |
| System jobs | Job name | Background work that no single Coworker or Knowledge Base owns, such as job description analysis, conversation compaction and automatic conversation review (quality-review) and the daily Value estimates (value-attribution). Rows are named by the job's recorded name, for example jd-analysis. Search, ingestion and image-reading jobs (knowledge-extraction, knowledge-extraction-test; see Knowledge extraction) are listed under Knowledge Bases instead. |
| Teams | Team | Each call is attributed through its Coworker's team, or else its Knowledge Base's team. Calls with neither land in Unassigned. |
| Users | Person | Calls made with a person's own API key, or made for them (for example an AI tool connected with their key, or a Coworker chat they started), count under that person. Calls that are not made for anyone, such as background jobs, land in Unassigned. |
Knowledge extraction#
When a file is added to a Knowledge Base, only the parts that need an image-reading model are
sent to one: scanned PDF pages, standalone images, and large pictures inside Word, PowerPoint and
OpenDocument files. Pages that already hold text are read directly and cost nothing. Each page or
picture sent is part of a model call recorded as knowledge-extraction, with its tokens and
cost. Making the embeddings that search uses is recorded separately, as ingestion.
What drives the cost is the number of pages and pictures read by the model, up to the Max model-read pages per document limit (see Knowledge limits). Pages are sent in groups of the size set by Page images per request, and a page that has been read is not read again if the work restarts. Rebuild search index never calls the image model, because it works from text that is already stored. Replacing a file reads it again.
Test image reading on the Image reading model panel makes one small call per test,
recorded as knowledge-extraction-test, shown in the Image-reading test row of the Knowledge Bases table and attributed to the administrator who ran it (see
Default models).
A Coworker, Knowledge Base, team or person that has since been deleted still appears, with "· removed" after its name, so past spend stays visible.
The Coworkers table is shown while Coworkers is on in Settings, then Features. When it is off, the table is left out and the totals still include every call.
Every call is counted once within a table. The tables are different views of the same spend, so do not add them together: a call appears under its Coworker, under that Coworker's team and under the person it was made for.
Usage and cost do not change when conversation content is removed under Conversation retention. Every total here stays the same.