Skip to main content
ModelHub can answer with any model your administrators have connected to the project. You either pick one yourself or leave the picker on Auto and let ModelHub choose per message.

​
The model picker

The picker sits in the composer and shows Auto or the name of the model you chose. Open it to see:
  • Auto at the top,
  • then the connected models grouped under Small, Mid and Large.
A dot marks the current choice. Each model appears once even if an administrator connected it twice. Your choice applies to every message until you change it. It resets to Auto when you reload the page or switch project. You can also change the model from an answer’s report card: the Use or Try suggestion chip selects that model for your next message.

​
What Auto does

With Auto selected, ModelHub reads each message before sending it and gives it a difficulty score between 0 and 1. The score decides which size of model the message needs: ModelHub then sends the message to the cheapest connected model of that size or bigger. Ordinary questions start below the Small line, so they go to the cheap model by default; something in the message has to argue for spending more. Signals that raise the score include code, several questions at once, and requests to prove, design, debug, refactor or explain step by step. Greetings, simple look-ups and translation requests lower it. When a score lands close to a boundary, a small, fast model is asked to classify the message and its answer overrides the score. This runs at the same time as the privacy screening, so it adds no waiting. The full scoring table, with every weight, is in How routing works.

​
Picking a model yourself

If you choose a specific model, ModelHub uses it. Routing never overrides an explicit choice. The message is still scored, so the report card can tell you whether your pick was Overkill or Underpowered for that question.

​
How models are sorted into sizes

Sizes are worked out from the model’s price, not from a fixed list:
  • Model names in the Claude Opus and Claude Fable families, and GPT-6 Astra, are always Large.
  • Names containing haiku, -mini, -nano, -lite or -luna are always Small.
  • Anything else is sorted by its average of input and output price per million tokens: up to 3.50∗∗isSmall,upto∗∗3.50** is Small, up to **8.50 is Mid, above that is Large.
Examples from CorePillar’s current catalogue: What you see in your picker depends on what your administrators connected.

​
What you see while it runs

While the answer streams, a line above the composer says where the message went: After the answer, the report card’s Details panel lists each signal that moved the score and the reason, such as “scored as a simple turn”, “classified as moderately complex” or “You picked this model”. See The chat screen.

​
Getting the most from Auto

  • Auto can only save money if the project has a spread of sizes. With a single model connected, every message goes to it.
  • If a report card says Underpowered and suggests a stronger model, click Try to use it for your follow-up.
  • If an answer looks thin on a hard question, pick a Large model for that message and switch back to Auto afterwards.

​
Savings

When Auto chose the model, ModelHub works out what the same answer would have cost on the most expensive model connected to the project, using the actual token counts. The difference is credited as Saved by auto-routing on the AI Usage page. Messages where you picked the model yourself are not counted.