How Grid routes

#1. Understand the request

Grid first determines what the request is asking the model to do and how difficult it is. It separately identifies requirements such as streaming, attachments, web access, tools, or other capabilities the request needs.

#2. Find models that can do it

Grid removes models that cannot handle the request or are not currently available. Only models that support the capabilities the request needs remain.

#3. Choose a model

Grid compares the remaining models on quality, cost, and availability, then selects one for the request.

The exact model can change as models, prices, and Elektric’s reviewed evidence change.

#4. Fall back when needed

If the selected route fails before output begins, Grid can try another eligible model. Once streaming output has started, Elektric does not silently restart the response with a different model.

If no eligible route succeeds, the request returns an error with its Request ID.

#5. Record

Elektric records the model and provider used, request status, latency, token usage, fallback activity, and provider cost where available. The Request ID connects the response to its Usage Log entry.

That means you can see what Grid chose without putting model-selection logic in your application.

#Examples

Grid chooses a suitable model for each request; that may not always be the model with the highest benchmark score.