Models are picked by what they are for — and Auto is the new default

Choosing a model used to mean knowing which one your deployment considers "large". Tiers were labelled S / M / L / XL, which told you the size and never the purpose — and sizes age, so this year's large quietly became next year's medium while the work stayed the same.

Tiers are now named for the job:

• Utility — cheap background micro-jobs you never wait for: naming a thread, writing an icon or a description, triage.
• Chat — the everyday round.
• Reasoning — multi-step analysis and planning, when the answer has to be thought through.
• Coding — writing and changing code, on the strongest model you have. Code is where a weaker model costs the most: a wrong patch is worse than a slow one.

They are ordinary nodes now, listed under Tiers in the AI menu next to Models and Providers — so you can rename one, re-rank them, rewrite what a tier is for, or add your own, and the edit survives every update.

Auto is the default selection for a new thread, and it now does something: it runs the tier the selected agent asked for, and your deployment's default when the agent asked for nothing. No extra call, no guessing — an agent that writes code lands on your coding model without anyone picking it.

Nothing you already had breaks. Agents that name the old heavy / standard / light tiers keep resolving to the same rung, and deployments configured through the old ModelTier:* settings keep their mapping until you label your models.