Monitor tenant usage
The Usage page shows how much your tenant uses the platform's models: how many tokens, how many requests, and how much audio. Use it to keep an eye on usage, find the models and API keys that use the most, and share the numbers with the teams behind them.
Only Owners can see the Usage page. See Roles in a tenant.
What gets counted
Every call to a model in your tenant is counted, whether it comes from Chat, an application, an agent or a RAG endpoint.
- Input - tokens sent to a model.
- Output - tokens a model sends back.
- Embedding - tokens used to create embeddings, for example when a RAG endpoint indexes your files.
- Rerank - tokens used to rerank search results.
- Total - all four token types added together.
- Audio duration - how much audio was sent for transcription. It's a length of time, not a token count.
- Requests - how many model calls were made. A call counts once, however many tokens it used.
Transcriptions count as Input and Output tokens, like chat, when the model reports them. A transcription model that replies in plain text reports no tokens. A successful transcription also records its Audio duration, which isn't part of Total. The built-in Whisper engine reports the duration for every format it returns; other models record it only when their reply includes a duration. Streamed transcriptions record no audio duration.
Usage through a model alias counts under the model the alias points to. Calls that can't be traced to an API key appear as No API key. Most of this is usage from Chat.
Anthropic's web searches aren't counted. When someone uses web search through the Messages API, Anthropic charges each search to the account behind the model's connector, and only the tokens show up here. Web fetch costs nothing beyond its tokens.
Codex CLI's web searches aren't counted either. When Codex searches the web on a gpt-6.x model that the platform calls at OpenAI directly, OpenAI charges each search to the account behind the model's connector. The search itself adds no tokens and no requests here. The model requests around it are counted as usual, including the search text that Codex passes to the model. See Web search on gpt-6.x models.
Open the Usage page
Select Usage in the tenant sidebar. The page opens on the last 2 hours and shows, from top to bottom:
- Token usage - a chart of tokens over time, and Window totals for the whole period.
- Token usage by model - a table with the exact numbers for each model.
- Audio duration - how much audio was transcribed over time.
- Requests - how many model calls were made over time.
Every card covers the same period. The page keeps updating while Live is on.
Look at a different period
Use the controls at the top of the page:
- Time range - pick a preset, from Last hour to Last 30 days, or Month to date.
- One month - see a whole calendar month. Choose it in Time range, then pick the month in Month. The list has the current month and the 11 before it.
- From and To - set your own start and end time.
- Drag across the chart - zoom in on a shorter window. Select Reset zoom to go back.
To answer "How much did we use this month?", set Time range to Month to date and read Total under Window totals.
Live keeps the page updating while the period ends at the current time. It turns off when you set To, zoom or look at a past month. Select Now to jump back to the current time and turn Live on again.
The page address keeps your settings, so you can bookmark a view or send it to another Owner.
See which models or API keys use the most
The Token usage chart shows how usage changes over the period.
- In Compare, select Models or API keys.
- In Chart models or Chart API keys, select the ones to draw. You can select up to 20, and the chart starts with the 20 that used the most.
- In Chart, switch between Bars and Lines. Bars show how each one adds to the total. Lines make it easier to compare them.
For example, to find out which team's application used the most this week, set Time range to Last 7 days and Compare to API keys. The biggest segment of each bar is the key that used the most in that interval.
Window totals always count every model and API key in the period, not only the ones on the chart.
Get exact numbers per model or API key
The table below the chart lists the exact token counts and audio duration for the period.
- In Group by, choose what each row is: Models, API keys, API keys, then models or Models, then API keys.
- To narrow the table, select models in Filter models or keys in Filter API keys. Select Clear filters to see everything again.
- Select a column heading, such as Total or Input, to sort by it.
For example, to see which models each API key used, set Group by to API keys, then models. Each key lists the models it used under it.
Filters change only the table. The chart and Window totals stay as they are.
See one API key's usage
Select API Keys in the tenant sidebar, select the key, then its Usage tab. See Checking usage.
Download usage per API key
Download a CSV file to share usage with the teams behind each API key or to analyze it in a spreadsheet.
- Set the period with Time range, or with From and To.
- Select Export API-key CSV.
Each row in the file is one API key and one model it used, with the tokens and audio for that pair. For example, if
your support-bot key used two models, the file has two rows for it.
The file only includes usage made with an API key. Chat conversations and other usage without a key are left out, so the file's totals can be lower than the totals on the Usage page.
For the full list of columns, see Export API key usage.
More resources
- Manage members and roles - who can see the Usage page
- API keys - create keys for each team or application, so their usage is counted separately
- Tenant metrics - usage across all tenants, for platform administrators