Dataset: AI Usage
Entity: AI Usage Measurement
Field ID: model_ref
Type: Text
Description: The normalised model reference, in <vendor>/<model> form, so the same model can be compared across providers. Usage the provider could not attribute to a model carries unattributed.
Note: Only populated on SURFACE_MODEL and LANGUAGE_MODEL records.
Source: Calculated
Transformation logic: The provider's raw model string is normalised to a vendor-qualified reference. The raw string is preserved in model_name.
App Mapping
GitHub Copilot | Calculated: normalised from the chat model dimension. |
Anthropic Claude (coming soon) |
|
Cursor (coming soon) | Calculated: normalised from the event |
OpenAI Platform (coming soon) |
|
OpenAI Codex (coming soon) | N/A - the Codex analytics API has no model dimension |
Reporting Use Cases
The Model Ref field is the comparable model identifier. Because the same model is named differently by each tool that serves it, this is the field to group by whenever a report spans more than one provider.
Model mix: Group by model ref with
SUM(output_tokens)orSUM(requests)to see which models the team actually relies on.Migration tracking: Charting model ref over time shows how quickly a team moves onto a newly released model, which is often the cheapest available efficiency win.
Filter on the breakdown first: Model attribution only exists on
SURFACE_MODELrecords. Without that filter the majority of records return an empty model.Interpreting unattributed: The provider reported usage without naming a model, or routed it automatically. The raw value is still in
model_name, andmodel_is_auto_routeddistinguishes automatic routing from genuinely unknown.