Google Antigravity rolls out Orchestrator for AI Ultra subscribers

Google is rolling out Orchestrator, a new feature inside its Antigravity AI coding platform, to Google AI Ultra subscribers. The pitch is simple: let the expensive model think and the fast model do the typing.

Under the hood, Gemini 4 Carbon handles planning and review, then hands coding and testing off to Gemini 4 Hydrogen. Google says the setup reduces quota use by 24% and completion time by 19%.

How Orchestrator splits the work

Carbon is built for high-level planning and for reviewing what gets produced. Hydrogen is tuned for rapid execution, which means writing code and running tests quickly.

Orchestrator, which launched on November 12, 2026, lets a single session move between those two modes on the fly. A developer does not have to manually switch models when a task shifts from reasoning about architecture to churning out functions. The system decides when to hand work down and when to pull it back up for review.

The benchmark numbers

Google’s early figures come from Terminal-Bench 4.1. On that benchmark, Orchestrator delivered success rates comparable to Carbon running on its own.

Orchestrator used 24% less quota and finished tasks 19% faster on average.

Who gets access, and at what price

Orchestrator is limited to Google AI Ultra subscribers. The Ultra tiers that include Carbon and Orchestrator are priced between $99.99 and $199.99 per month.

The Pro subscription costs $19.99 per month. It includes only the Argon and Hydrogen models, so Pro users do not get Carbon or Orchestrator.

Hydrogen itself is being made available gradually to both Pro and Ultra subscribers.

Antigravity supports Orchestrator across its desktop apps, command line interface, and IDE extensions.

A busy week for Gemini 4

Orchestrator did not arrive in isolation. Google launched Gemini 4 Carbon on November 10, just two days earlier.

That Carbon launch came alongside a preview of /swarm, a multi-agent feature.

Antigravity first launched on its own in May 2026 and has since grown to offer a set of specialized models aimed at different parts of engineering work.

What this means for developers and the AI coding race

The most practical takeaway concerns cost per task. A feature that trims quota consumption by 24% while holding success rates steady effectively stretches what Ultra subscribers get for the same monthly fee.

The broader trend here is model routing. Instead of asking one giant model to do everything, platforms are starting to match each subtask to the model best suited for it. Orchestrator is Google’s version of that approach baked directly into a single session.

The /swarm preview is another thread to follow. For now, Ultra subscribers get the first look, and Pro users get Hydrogen as a preview of the execution half of the equation.