Auto model routing: the right model tier for every task

Most requests don't need the most expensive model, and some need nothing less. Auto now makes that call per run. It sits at the top of the model picker with three tiers (Fast, Balanced, and Max), and new members start on Auto everywhere. Members who already chose a default keep it.

How it works

• Fixed surfaces route directly. Building or editing a workflow, app, or skill always runs on Max; agent workflow executions default to Balanced.
• Chat is classified. Your workspace's utility model sorts each message into build, execute, or answer, mapped to Max, Balanced, and Fast. A question about your own data counts as execute.
• Fails up, never stops. If classification fails, the run goes to Max. Routing can never block a run.
• Sticky while warm. A conversation stays on its tier for five minutes so follow-ups keep hitting the same prompt cache, and the choice is stored on the run so resumes stay consistent.
• Outage-tolerant. Each tier has a fallback list spanning three providers.

The classifier scored 50 out of 50 across three runs on an eval built from the real (anonymized) distribution of chat openers.

Also shipped

• Grok 4.5, and the GPT-5.6 family (Sol, Terra, Luna) with 1M-token context windows.