Introduction
Model records behave like agent records with two differences that account for most of the confusion: a model is measured in spans rather than traces, and it has nothing to instrument.
A model shows no spans
The empty state says spans appear once the model starts processing requests. Several different situations produce it, and they need different responses:
Situation | How to tell | Do |
|---|---|---|
The platform does not collect usage for it | It is not an OpenAI model | Nothing to fix on your side. Usage collection for models exists for OpenAI today |
Its completions are not stored | An OpenAI model whose completions your application does not ask OpenAI to store | Store the completions with OpenAI so the platform can collect them |
The connection is not healthy | The connection the model was imported through is inactive or failing | Fix the connection |
Nothing calls it | No application uses this model | Nothing — it is inventory, not a fault |
Note: A model's spans come from its provider, not its callers. The platform collects a model's usage from the provider, so instrumenting an agent does not fill a model's spans. Check the model's platform, whether its completions are stored, and its connection before you look at agents.
A model is missing from the import list
The list is the connection's answer to "what can you serve", asked at that moment. If a model is not in it, the connection cannot see it — its credentials are scoped more narrowly than you expect, or the platform does not expose that model to this account.
"This connection returned no importable models" means the connection is working and has nothing to offer. That is a different problem from a connection that fails.
Model cost looks wrong
Are you looking for it in the agent total? — Spans that belong to a model asset are counted only in the Model view, not in the Agent view — see Agent View vs Model View.
Is there a rate for this model? — Usage is observed either way; cost needs a rate. A model with none shows activity and no spend.
Are token counts reported or derived? — Derived counts are an approximation, and accuracy varies by integration rather than by model.
Are you comparing to an invoice? — They answer different questions. See How Cost & Tokens Are Calculated.