Keep track of usage and costs

The Usage page shows what AI calls cost, and what for. An AI call is one request from SubKit to an AI model. Every call costs money. At the top, one sentence sums up the time window, for example "This week: $84.20 for 1,240 AI calls." When costs rise, the sentence names the largest driver.

What you need #

• To view the page: stats:read. By default only the Admin role has it.
• For Person and Conversation under Break down by: stats:read:all. Without it you do not see these tabs.
• To change the budget and prices: settings:write. The Admin role has both.

Choose the time window #

Choose the Time window at the top right: 24 h, This week, Last week, This month or Last month. Weeks start on Monday, and your time zone applies. All figures compare the window with the same span before.

How to read the trend:

• ↑ 18% or ↓ 18%: more or less than in the window before.
• ≈ same: less than 5% difference.
• New: not used in the window before.
• –: there is no data for the window before.

The five tiles #

• Cost: the estimated cost of all AI calls in the window.
• AI calls: how often SubKit asked an AI model. When calls failed, it shows for example 3 failed. A click opens the Failed AI calls.
• Budget · Month: how much of the monthly budget is used. This tile always covers the current month.
• Saved by cache: what reused input saved compared with full price.
• From the cache: the share of input that was reused instead of paid again. Below it, Cache per flow shows the share for each source. Higher is cheaper and faster. No data means the provider reports no cache for that flow.

The forecast sits below the budget. SubKit projects it from the last 7 days:

• Lasts to month end.
• About 94% by month end. The budget is getting tight.
• Used up on the 26. The budget does not last to month end.
• Used up. Chat, creating and fixes are paused.
• No forecast yet. Too few days this month. This applies to the first two days of a month.

When a model has no price, the cost shows at least in front. The amount is then a floor. The real cost is higher. Below it you see how many calls have no price, for example Calls without a price: 12. When no prices are known at all, the tile shows Cost unknown.

Worth a look #

Worth a look lists what changed clearly or costs money. Each notice has a matching action:

When nothing stands out, it says "Nothing unusual in this window."

Cost by origin #

Cost by origin shows the cost per day, or per hour for 24 h. The bars are stacked by origin:

• Chat: replies in conversations.
• Designing workflows: creating and fixing workflows in chat.
• Summaries: short summaries of runs and content.
• Checks: AI checks drafts and results.
• Knowledge & files: reading and describing sources and files.
• Workflow runs: "Ask AI" steps in runs.
• Other: everything without a known origin.

As table shows the same values as a table with a Total column. As chart switches back. Click an origin in the list below to open its details. When prices are missing, the chart shows token amounts instead of cost.

Break down by #

Under Break down by you choose Workflow, Person, Model or Conversation. Each row shows Cost, AI calls, Share and Trend. Some tabs add a column: Runs for Workflow, Error rate for Model, Last active for Conversation.

The last row collects what belongs to no row:

• No workflow (chat, designing, knowledge …)
• No person (automatic): schedules, webhooks and background work such as cataloguing
• Other for models
• No conversation (runs, knowledge, checks …)

That way all rows together add up to the total at the top.

Open details #

Click a row to open its details at the side:

• Workflow: Cost per run, AI calls per run, Most expensive runs and Models used. Open workflow takes you to the workflow.
• Person: Where usage comes from, plus their Workflows and Conversations.
• Model: Share from cache, Error rate, Price per 1M tokens and Response time (typical).
• Conversation: cost and AI calls. Open conversation takes you to the chat.
• Origin: under Largest items, the most expensive workflows, people or models of this origin.

Long conversations get more expensive because every reply reads the history so far. A new topic is worth a new conversation.

You open Failed AI calls from the AI calls tile. Failed calls still cost money. The sheet names the amount. Under Where they happen you see the workflow or conversation and the reason: Provider error, Input too large or Budget used up.

Adjust the budget #

The monthly budget is the maximum number of tokens per month.

1. Click Adjust budget under Worth a look, or the forecast in the tile. Without a budget, choose Set a budget in the tile.
2. Enter a value under Tokens per month.
3. Save. The message Budget saved confirms the change.

You need settings:write for this, so by default the Admin role. Without it you only see the current state and "Ask someone with the Admin role."

What happens as tokens are used:

• From 75% a notice appears.
• At 100% chat, creating and fixes pause until the first of the month.
• Approved workflows keep running.

Without a budget, AI calls are unlimited.

Set a price #

When a model has no price, its calls are missing from the cost.

1. Open the model, for example through View model or the Model tab.
2. Choose Set a price.
3. Enter Input per 1M tokens and Output per 1M tokens and save.

The message Price saved confirms the change. The page then calculates with the new price. You need settings:write for this too, so by default the Admin role. The price applies to your workspace. In the settings it sits under tokens.prices.

Export #

1. Choose Export at the top.
2. Confirm with Export.

You get a CSV file for the chosen window. The file has one line per day, origin and model, or per hour for 24 h. The columns are the time, Origin, Model, AI calls, the four token kinds, Cost and Currency.

Technical details #

Expand Technical details at the end of the page. It shows:

• Input tokens, Output tokens, Cache read, Cache written and Total.
• Models without a price and the number of calls without a price.
• The Budget in tokens, for example "120,000 of 150,000 tokens · 30,000 left".
• The Price source: built-in list prices per model, or the prices set in tokens.prices.
• Your Time zone and the Internal flows per origin.
• The Query that fetches the same numbers through the CLI or the API.

The same numbers through the CLI or the API #

sk stats llm and GET /v1/stats/llm return the same numbers as the page. Each answer includes tokens, cost and the savings from the cache.

sk stats llm --bucket day --from <from> --to <to> --group-by workflow

• group_by breaks down by model, flow, user, workflow, conversation or execution. At most two values at once.
• result= filters by result: success, quota_exhausted, context_too_large or error.
• workflow_id= shows only the calls of one workflow.
• Breakdowns by user or conversation need stats:read:all. That does not apply when user_id is your own ID.

When something does not work #