Antigravity’s Orchestrator lets Carbon and Hydrogen share one coding session

Google is rolling out Orchestrator in Antigravity, its agentic coding environment, to Google AI Ultra subscribers. According to the announcement, it lets one session run multiple models instead of routing the entire task to a single model.

One session, several models at work

Google’s post, published on its official X account, says Orchestrator is rolling out in Antigravity to AI Ultra subscribers. Techmeme’s summary of the news adds that it can switch dynamically between models such as Carbon and Hydrogen within the same session. Until now, a task in a coding agent was typically tied to the one model the user picked at the start. Orchestrator changes that by letting the tool decide, as work proceeds, which model handles which part.

The announcement is short. It names AI Ultra as the plan that gets the feature first. It does not say when other plans might follow, which regions are covered, or how the rollout is staged.

Why Hydrogen is the natural partner

Hydrogen launched the same day. Google describes it as built for rapid execution across diverse agentic workflows, and it is available to paid AI Pro and Ultra users in Antigravity, according to 9to5Google’s report as listed by Techmeme. Our report on the Hydrogen launch covers its pricing and benchmark position against Argon.

That context shows why mixing models makes sense. A fast, lower-priced model suits routine steps, while a heavier model can be kept for the hard parts of a task. Orchestrator appears to be the mechanism that makes that split automatic, though Google’s post does not spell out the routing logic.

Who this affects in practice

For developers on AI Ultra, the practical change is that a long session no longer forces a single choice between speed and depth. They can leave the choice to the tool, at least in principle. For AI Pro subscribers, Hydrogen is already in Antigravity, but Orchestrator is not mentioned for their plan, so the mixing feature is not yet available to them according to the post.

For teams comparing coding agents, the move follows a broader pattern. Rivals such as Anthropic’s Claude are expanding what a single assistant does in one workspace, and Google is now doing so by combining its own models rather than relying on one.

What the announcement leaves open

Several details are missing from the material available at the time of writing:

• How Orchestrator chooses a model for each step, and whether users can override or configure that choice.
• How usage limits or quota are counted when more than one model works in a session.
• Whether models from other vendors can take part, or only Google’s own.
• Whether any speed or cost benefit holds up on real projects. Google has published no figures for this feature in the post we read, so no claim can be checked.

Techmeme, which surfaced the story, links Google’s post as its source. Google’s own documentation for Antigravity should clarify the rollout details as they are published. Until then, the verified facts are limited to the feature’s name, its first audience, and the idea behind it: one session, several models.