Introduction
The same spend, sliced two ways. Agent view attributes cost to the thing doing the work; model view attributes it to the thing being called. Both totals describe the same money, which is exactly why you must never add them together.
Where to find it
AI Governance > Cost Explorer. The Agent / Model toggle sits under the title.
What each view is for
Agent view | Model view | |
Attributes cost to | The agent that ran | The model that was called |
Answers | "What is this capability costing us?" | "What are we spending on this provider?" |
The owner is | A team, with a budget | Usually nobody in particular |
Action it supports | Fix or fund a specific workload | Negotiate, consolidate, or switch |
Unit underneath | Traces | Spans |
Note: These are two lenses on one pool of money, not two pools. A model call inside an agent's run is counted in the agent view and in the model view. Adding an agent total to a model total double-counts every call. If a report shows both, it must present them as alternative breakdowns of the same figure.
When agent view is the right lens
Agent view is the accountability lens, because an agent has an owner and a purpose. It is what you need for:
Chargeback and budgeting — An agent maps to a team and a use case. A model does not.
Justifying a workload — ROI compares an agent's cost against the outcomes it produced. Only the agent side of that comparison exists in model view — see Reading the ROI View.
Finding the expensive workload — One agent consuming a disproportionate share is a finding you can act on, because someone owns it.
When model view is the right lens
Model view is the procurement lens. It aggregates across every agent, which is exactly what agent view cannot do:
Understanding provider exposure — Spend spread thinly across twenty agents is invisible per agent and obvious per model.
Seeing the effect of a model choice — The cost difference between two models is a property of the models, and it shows up here first.
Finding shared infrastructure nobody owns — A high-spend model that no single agent dominates is shared, and shared and unowned is the most common cost blind spot.
"Agent view tells you who to talk to. Model view tells you what to negotiate."
Reading them against each other
The interesting findings live in the gap between the two views, and they are only visible if you look at both:
Pattern | Likely meaning |
One model dominates, no single agent does | A default model used everywhere. A change here moves everything at once |
One agent dominates, spread across models | A complex workload calling several models per run. Look at the trace shape |
An expensive model with one agent behind it | The cleanest case — one owner, one decision |
Model spend far exceeds the agents you recognise | Something is calling that model outside the agents you have registered |
That last row is the one worth acting on. It is a coverage gap wearing a cost report's clothes, and the fix is in the registry rather than in Cost Explorer.
Group by, and the honest limit
Within either view, grouping narrows the picture further — Agent, Owner or Cost Type in the Agent view; Model, Provider or Owner in the Model view. Grouping does not change the total; it changes how the same total is divided.
One limit worth stating in both views: spend that is not attributed to anything still exists. Any breakdown covers the spend it can attribute, and the remainder is real money that no slice of this screen will show you. It is the same caveat that makes a return-on-spend figure only as good as its measured spend — see Reading the ROI View.