Picking a model by name is a chore that never ends. A cheaper one appears and every node that named the old one is now overpaying. A provider gets busy for ten minutes and a run that would otherwise have been fine fails.
Both of those are now handled for you.
Categories
Agent, Guardrails, Gmail Send and the XLSX nodes, along with agent phases, can be set to a category instead of a model:
- Fast for classification, routing and short extraction.
- Balanced for most real work.
- Advanced for the steps where quality is the whole point.
Within a category the platform prefers the cheapest model that can actually do the step, then the quicker one, then the better one. When we add a stronger or cheaper model to the catalogue, your nodes pick it up without you editing anything. Categories describe how hard the task is, not a latency guarantee.
Nodes that name a specific model keep naming it. Nothing was migrated underneath you.
Narrow the pool when you want to
A category is a managed pool, and sometimes you want a smaller one. Open Customize models on any node with a category and you get the list, sorted by price, with a live credit range and the count of what you have selected. Untick what you would rather not pay for.
Your selection is a snapshot, which matters more than it sounds: a model added to the catalogue next month does not silently join a list you curated. If you want it, you say so.
Fallbacks
Every model on the platform has a fallback configured behind it, and a failed call walks that chain, up to five in a row, before the run gives up. The chain respects what the step actually needs, so a step that sends an image only falls back to a model that can read images, and it never leaves your category or steps outside a list you narrowed.
Voice calls fail over too, mid-call, rather than dropping the caller because a provider had a bad minute.
If the chain runs out, the run fails and says which pool was exhausted. That is deliberate. Quietly finishing a run on a model you deselected would be the worse outcome.
What to do with it
Open a workflow, set your agent nodes to a category and look at the credit range that appears. For most teams the interesting discovery is how much of a workflow was running on an expensive model out of habit rather than need.