Models & Thinking

Which model runs your requests, how hard it thinks, and what to pick when.

The model picker

The picker sits in the chat input. It lists every model Airbuild can run, grouped by provider, and each row shows three relative ratings — cost, speed and quality — plus a one-line note on what that model is good for.

Your choice is saved per project and applies to every chat in it. Every model bills into the same AI Tasks, so switching changes what a request costs in tasks, not how you are billed.

Claude Sonnet 5 is the default. If you never open the picker, that is what runs.

What you can pick

A rough rule. The top of the list — Fable 5, Opus 5.5, GPT-6 Astra — earns its price on architecture, stubborn bugs and large refactors. The middle — Sonnet 5, GPT-6 Sol, GPT-5.6 Terra, Gemini 3.1 Pro — handles ordinary feature work. The fast ones — Haiku 4.5, GPT-6 Luna, Gemini 3.7 Flash, DeepSeek V4.1 Flash, Mistral Medium 3.5 — are for small, well-described changes: a label, a colour, one more field.

How hard the model thinks

Beside the picker is the effort control. It sets how much reasoning the model may spend before it answers:

• Adaptive — no ceiling; the model decides step by step how much to think
• Low — fast and cheap
• Medium — balanced, cheaper and faster than High
• High — the model's maximum, slower and pricier

Adaptive exists only for the Claude 5 family — Fable 5, Opus 5.5 and Sonnet 5 — which can vary their own thinking from step to step. For those three it is also the default, so a simple step stays cheap and a hard one gets the room it needs.

Every other model offers Low, Medium and High, and starts at High — its real ceiling. That is deliberate: reasoning is billed as output, so High is the expensive end of a model, and you dial it down when a task does not need it.

A model whose provider does not vary its reasoning shows no control at all. A switch that does nothing is worse than no switch.

A project stays on the model it was set to

A project pinned to a particular model keeps running on it, priced as before, even after a newer model takes its place in the picker. Nothing is switched under you. To move a project forward, open the picker and choose the new model yourself.

Model policy in an organisation

An organisation can set the rules for everyone in it, in its settings:

• Allowed models — restrict the picker to a chosen set, so every project in the organisation stays inside an agreed range
• Default model — what a new project starts on; it has to be one of the allowed ones

Members keep the picker for everything still allowed. See Team & Roles.

What it costs

Every model bills into the same AI Tasks — see Plans & AI Tasks. A pricier model and a higher effort both consume more per request, which is the whole reason both controls sit next to each other.