Orchestrator Antigravity: how Google lets Carbon (or Argon) plan and Hydrogen build

Google launched Orchestrator Antigravity, a new feature of its AI-powered development environment, on November 12, 2026. The feature arrives alongside Gemini 4 Hydrogen and changes how developers work with the Gemini 4 models: instead of picking a single model for the whole task, you select Orchestrator and the system switches between two models inside the same conversation — a frontier model to plan and Hydrogen to execute.

What is Orchestrator

Orchestrator is an option in Antigravity's model selector, not a separate product or a new model. It appears as Orchestrator (Carbon) or Orchestrator (Argon), and the name in parentheses tells you which model does the planning. In both cases, Hydrogen does the execution.

  • Orchestrator (Carbon): planning with Gemini 4 Carbon, Google's strongest model for long-horizon tasks, architecture and cybersecurity.
  • Orchestrator (Argon): planning with Gemini 4 Argon, the frontier model launched in September, the same one behind the agents that beat Minecraft on their own.

Hydrogen is the execution model: fast and efficient, built to write code, run tests and verify changes. According to Google, it completes tasks 2.5 times faster than Gemini 3.8 Flash at its default setting and costs $1.50 per million input tokens and $7.50 per million output tokens via the API, with a 1-million-token context window.

How it works

Orchestrator works in a single conversation, with a single agent, switching models as the work evolves:

  1. Read-only planning: with each message, the planner (Carbon or Argon) starts by exploring the code — reading files, mapping dependencies, resolving ambiguities — without being able to edit anything. It then produces a concrete plan with explicit re-planning triggers, the conditions that should make the plan be revised.
  2. Execution with Hydrogen: control passes to Hydrogen, which applies the edits, updates packages, runs builds and tests, and checks the result against the plan.
  3. Automatic handback: if Hydrogen hits friction — repeatedly failing tests, a build error, unexpected behavior — control returns to the planner, which inspects the state, repairs the plan and resumes execution.

What Orchestrator is not

Google stresses one point: Orchestrator is not a static router that classifies the prompt up front and sends the whole task to a single model. The switch happens dynamically and bidirectionally during the task, inside a single response, with no manual model switching.

One practical detail: when a task needs no changes — a question, a review or an explanation — the planner can answer directly, without calling Hydrogen.

How to turn it on and use it

To use Orchestrator Antigravity, open the model selector below the prompt box in the Antigravity app and choose Orchestrator (Carbon) or Orchestrator (Argon). The choice applies to the whole conversation until another model is selected. It also works in the agy CLI (/model command) and in Antigravity's IDE extensions.

For programmatic control, the Antigravity SDK lets you build explicit multi-model pipelines — for example, Hydrogen agents for issue triage and coding, and a Carbon agent for larger refactors and security checks.

Who can use it

  • Orchestrator (Carbon) and Orchestrator (Argon): Google AI Ultra subscribers.
  • Orchestrator (Argon): also for paid Google AI Pro subscribers.
  • Enterprises: both via Gemini for business.
  • Standalone Hydrogen in Antigravity is for paid Pro and Ultra subscribers; outside Antigravity, Hydrogen is already in the Gemini API, AI Studio, the Gemini app and Google Search.

Google says more models and plans will be supported in the future and that the rollout is gradual — it can take a few days to reach every user. One discrepancy is worth noting: Antigravity's documentation model table currently lists Orchestrator for Ultra only, while the official launch post includes Orchestrator (Argon) for paid Pro. When in doubt, check availability in the model selector itself.

The numbers: what Google promises

On Terminal-Bench, Orchestrator reached near-Carbon task success while consuming on average 24% less quota and finishing 19% faster, according to Google. One detail: the DeepMind post cites Terminal-Bench 4.0, while the Antigravity blog cites 4.1 — the percentages are the same, but the benchmark version differs between the two official posts. And all figures are Google's own measurements, with no independent replication so far.

Standalone Hydrogen

For standalone Hydrogen, Google cites 51 points on the Artificial Analysis Intelligence Index at the default high setting (54 at the xhigh level), above Argon's 50, and 57% on Terminal-Bench 4.1, versus 18.2% for Gemini 3.8 Flash.

Usage tips

  1. Use Orchestrator for long tasks: new features, large refactors and complex debugging, where good planning saves rework.
  2. For short, well-defined tasks, standalone Hydrogen is usually enough — Google itself recommends the single model when no planner is needed.
  3. State the goal, constraints and acceptance criteria in the first message: the better the context, the better the Carbon or Argon plan.
  4. Read the plan before execution: it carries the re-planning triggers — knowing what would change the plan helps you follow the work.
  5. Without Ultra, Orchestrator (Argon) on Pro delivers the same Hydrogen execution, with Argon planning.
  6. Combine it with /swarm for parallel work: Carbon coordinates while Hydrogen acts as worker and verifier in isolated workspaces, according to Google.
  7. If execution stalls repeatedly, let the handback happen: the planner revises the plan instead of forcing fixes.
  8. Remember that the reasoning level is fixed in Orchestrator — you cannot tune effort as with standalone Hydrogen, which has four levels (low, medium, high and xhigh).

Pros and cons

Pros

  • One prompt, no manual model switching: frontier planning and fast execution in the same conversation.
  • Quota savings: 24% less quota and 19% faster than Carbon alone, according to Google.
  • A plan with re-planning triggers and automatic handback when execution hits friction.
  • Orchestrator (Argon) available on Pro, without requiring Ultra.
  • Works in the app, the CLI and IDE extensions, plus the SDK for programmatic pipelines.

Cons

  • Paid plans only: no Orchestrator on the free tier.
  • No control over the reasoning level: the configuration is fixed in Orchestrator mode.
  • No per-step log of which model performed each action, according to the current documentation.
  • The 24% and 19% are Google's own averages on a single benchmark — gains may vary on real repositories.
  • The saving is Antigravity plan quota (metered by work performed), not a direct dollar saving.
  • Gradual rollout: it may not have reached every subscriber yet.

Why it matters

Orchestrator Antigravity solves a real dilemma in AI-assisted development: using the strongest model for everything burns quota and time; using only the fast model loses planning quality. By splitting the work — frontier model to think, Hydrogen to do — Google aims to deliver near-Carbon quality at lower cost and latency. If the numbers hold up in real use, it is the kind of feature that changes the cost of running agents on long tasks.

Also on Geekiki

Sources

Image credit: Google / Divulgação

Transparency note: this article was prepared with the assistance of artificial intelligence, based on the official sources and specialist coverage cited at the end, and reviewed before publication.

Por Muse