Every AI feature in wxrks — Context Sensitive Translate, Translation Smells, Fix Tags and the rest of wxrks Augmented Actions — consumes tokens, the unit AI providers bill by. The Usage tab in Model Management shows how many tokens your account has consumed, which AI features consumed them, and which parts of your organization they came from. This article explains how to read it and how the numbers add up.

The tab loads your account's consumption for the current calendar month.

⚠️ Warning: Do not confuse this tab with Account Settings → Usage in the sidebar. That page reports your plan's consumption, such as processed words — see Processed Words and wxrks Limits. AI tokens are reported only here, on the Usage tab inside Model Management.

The buttons at the top of the tab set the window every number below them is calculated over:

All windows are calculated in UTC, so a call made late in your local evening may land in the following day's or month's figures.

ℹ️ Note: Changing the period reloads the chart and both breakdown tables, and the Tokens tile — but not the other two tiles, as described next.

⚠️ Warning: Forecast (month) and Tokens YTD do not change when you switch periods. Selecting 90 days and then comparing the Tokens tile against Forecast (month) compares a 90-day window against a one-month projection. Only the Tokens tile answers "how much did we use in the period I selected".

wxrks fits a straight line through this month's daily consumption so far, extends it across the days remaining in the month, and adds that to what you have already used. Two behaviors are worth knowing:

💡 Tip: Treat the forecast as a trend indicator, not a commitment. A single large project pushes it up steeply, and it has no knowledge of your project pipeline.

The chart plots consumption over the selected period, split into Input and Output.

ℹ️ Note: The chart stacks only Input and Output, which is why the height of each bar equals that bucket's real total and reconciles with the Tokens tile above. Thinking tokens are shown in the tooltip rather than as a third stacked band, because adding them as a band would count them twice.

Two things about the axis are easy to misread:

If the account made no AI calls in the selected period, the chart area shows No model settings yet. instead of a graph.

Two tables sit below the chart, both covering the selected period and both sorted by Total descending, so your heaviest consumers appear first.

Shows which AI feature consumed the tokens, using the same feature names as the rest of Model Management — CS Translation, Smells (Translation Smells), Spell Check, and so on.

ℹ️ Note: The same feature name can legitimately appear on two rows. wxrks has recorded feature names in more than one format over time, and the table groups by the stored value while displaying the current name. Both rows are real; add them together for that feature's true total.

Shows which part of your organization the consumption came from, which is what you need for internal cost attribution.

This is the part that most often looks wrong, so it is worth being precise.

The rule: Input + Output = Total. Thinking is already inside Output, not a number you add on top.

Take the CS Translation row from the capture above:

The reflex is to add all three visible figures — 524.3K + 195.9K + 117.7K = 837.9K — which overshoots the Total of 720.3K and makes the table look like it miscounts. It does not. The correct sum is 524.3K + 195.9K = 720.2K, which is the reported total give or take one rounding step. The 117.7K thinking tokens are a breakdown of the 195.9K in Output, not an addition to it.

💡 Tip: The tables abbreviate large numbers (720.3K, 1.8M), so a hand-check can land a rounding step away from the displayed total, as it does above. The underlying figures reconcile exactly.

wxrks reports Total exactly as the AI provider reported it, and derives Output as every token the provider did not count as input. That definition is used because AI providers do not agree on where thinking tokens belong:

Because a single row can aggregate calls across several providers, no fixed piece of arithmetic involving the thinking figure would be correct for every row. Deriving Output from the total is the one rule that holds everywhere — and it is the same rule wxrks bills on, so the Total column is the figure to reconcile an invoice against.

ℹ️ Note: A missing incl. line does not mean the model did no reasoning. It means the provider did not report a separate count — Anthropic and Azure OpenAI never do.

The Model Management tab lists the model configurations serving your account and reports each one's consumption in its Usage column, so you can see which configuration is actually doing the work. This includes the wxrks-managed default configurations that ship with your account, not only configurations you created yourself.

So ↑ + ↓ is that configuration's provider-reported total, exactly as in the Usage tab. A row showing — recorded no calls in the selected period.

ℹ️ Note: The Usage column belongs to the Advanced Model Management list, which requires the FUSION plan or higher. On other plans the list is replaced by an upgrade prompt, and the Usage tab remains your view of consumption.

💡 Tip: A configuration you created that shows — month after month is either unreachable or redundant. Check its Scope, Operations and enabled state against the routing rules before assuming the feature is unused.

"Input + Output + Thinking doesn't equal Total." It is not supposed to. Thinking is already counted inside the total — see How the numbers add up above. Add only Input and Output.

"My hand-check is off by a few hundred tokens." The tables abbreviate to one decimal place. 524.3K + 195.9K displays as 720.2K against a reported total of 720.3K; both describe the same underlying figures.

"The Tokens tile and the Forecast tile disagree." They are measuring different windows. The forecast always covers the current calendar month, whatever period you selected. Compare like with like by selecting Current Month.

"The same feature appears twice in Tokens by operation." Expected, and both rows are real. wxrks has stored feature names in more than one format over time; add the rows together.

"My tables are empty but I know the AI features are being used." Check the period first — a rolling 30 days window excludes anything older. If Year to date is also empty, confirm the AI features are actually enabled for the org units in question, as described in wxrks Augmented Actions.

"Yesterday's usage is missing." Periods are calculated in UTC. Calls made late in your local day can fall into the next UTC day, so check the adjacent bucket in the Daily chart. Note also that days with no calls are skipped by the axis entirely.

"I don't see a Usage column on the Model Management tab." The column belongs to the Advanced Model Management list, which requires the FUSION plan or higher. The Usage tab itself is available on every plan.