Airbuild 1.5: Sonnet 5.5, Haiku 5.5 and Mistral Large 4

Version 1.4 was about work that does not stop. Version 1.5 is about the decisions around that work: which model you are spending on, which of your endpoints are open because you meant them to be, and what a project knows about your standards before the first prompt.

Sonnet 5.5 and Haiku 5.5 are the new defaults

Claude Sonnet 5.5 takes over as the default model and Claude Haiku 5.5 as the fast one beside it. Mistral Large 4 replaces Mistral Medium 3.5 in the picker. Haiku 5.5 costs a tenth of what Haiku 4.5 did per token and reads far more at once.

Reasoning effort also gets its own control. Every row in the model list used to carry four letters — A, L, M, H — which made the list wider and explained nothing. It is a chip next to the model picker now, showing the active model’s level as an icon and a word.

Your data outlives the address

Unpublishing used to do two opposite things at once: the logs and the custom domain went immediately, the database stayed forever. Neither is what anyone expects.

Now everything stays for 30 days — database, logs and custom domain. Only the public address is dropped. The same rule applies when a plan change takes extra deployment targets offline: they stop serving and stop billing, and their data is kept for the same 30 days.

An endpoint that is open on purpose

Before a publish, the gate lists every endpoint reachable without a login. Until now there were two honest answers: secure them, or publish anyway.

There is a third now. Tick the ones that belong to the public — a webhook, a health check, a price list — and the AI marks them as intentional, so the next publish does not stop on them again.

Every deployment target gets its own database

An additional target — a staging environment next to production, say — now gets its own database the way the default target always had one, and keeps it for as long as the target exists: created with the target, kept across every publish, removed with the target.

Project rules, before the first prompt

You can hand a project a rules file while you are creating it: a Markdown or text file up to 100 KB saying how this project is to be built — coding standards, naming, things to stay away from. The rules apply to every AI chat in the project from the first turn, not from the point where you remember to mention them.

A start screen that does one thing

The entry screen asked what you want to build and listed everything you had already built, both at once — whichever you came for, the other was in the way. Home is about starting now: the composer and the five projects you last touched. Everything else lives in the project list.

The AI Act notice moved with it. Article 50(5) asks for the notice to be clear and legible no later than the first interaction; ours had drifted to the foot of the start screen, below the project tiles and so below the fold on an ordinary display.

Published sites now live on airbuild.space

A site published without a domain of your own used to get an address at xaio.app. New publishes get one at airbuild.space, and the links stored for sites that already existed moved along with them. The old addresses keep working.

Everything else

The full release notes live on the Version 1.5 page.

• Five test runs in one message collapse into one line from the second run on.
• A declined top-up no longer blocks the next one — the open invoice is settled instead of guarding against itself for up to two days.
• Duplicating a large project no longer dies at the gateway’s two-minute timeout and leaves the copy orphaned.
• An amber note says which areas are being updated while backend services are replaced. It warns, it does not block.
• Logs, the code viewer and the generated API documentation follow your theme instead of staying dark.
• A preview cannot listen in on the workspace: messages are addressed to its exact origin and checked on arrival.
• The database view shows how many tables a project has, and how many match an active filter.
• The spreadsheet suggestion asks for an app, not a table — formulas become logic instead of stored results.
• The chat reports connected integrations by name again, and a code search that finds nothing is a normal result, not an error.