Antigravity Orchestrator: Power-User Guide to Carbon, Argon and Hydrogen

Last updated: November 16, 2026

Orchestrator is a model selector option in Google Antigravity, launched November 12, 2026. It pairs a planner, Gemini 4 Carbon or Gemini 4 Argon, with the faster Gemini 4 Hydrogen, which does the edits and tests. Google AI Ultra gets both versions, and paid Google AI Pro gets Orchestrator (Argon). You cannot set its thinking level.

Contents

What Orchestrator is

Google introduced Orchestrator alongside Gemini 4 Hydrogen in a Google Antigravity blog post on November 12, 2026. It appears in the model selector of the Antigravity App, the Antigravity CLI and the Antigravity IDE, as listed on the Antigravity models page.

Instead of running one model for a whole task, Orchestrator splits the work by role. A larger planner model explores the code, writes the plan and reviews progress. Gemini 4 Hydrogen, Google’s fast, lightweight Gemini 4 model, applies file edits, writes code, and runs builds and tests.

There are two options, and the one you pick sets the planner:

Model selector option Planner Implementation Available on
Orchestrator (Carbon) Gemini 4 Carbon Gemini 4 Hydrogen Google AI Ultra
Orchestrator (Argon) Gemini 4 Argon Gemini 4 Hydrogen Google AI Ultra; Google AI Pro (paid plans, or trials with eligible premium devices)

For background on the two planners, see our guides to Gemini 4 Carbon and Gemini 4 Argon.

Plan availability

The models page lists which reasoning models each plan can select. For the three Gemini 4 models and the two Orchestrator options, it shows:

Model Free and Google AI Plus Google AI Pro Google AI Ultra Enterprise
Gemini 4 Carbon No No Yes Yes
Gemini 4 Argon No Yes (paid or eligible trial) Yes Yes
Gemini 4 Hydrogen Yes Yes Yes Yes
Orchestrator (Carbon) No No Yes No
Orchestrator (Argon) No Yes (paid or eligible trial) Yes No

The launch post said Hydrogen was rolling out to paid Google AI Pro and Ultra subscribers, while the current models page and plans page list Hydrogen for all plans. The launch post also says enterprises can access Hydrogen and Orchestrator in Gemini for business, although the Enterprise column of the Antigravity table does not list Orchestrator. Google says more models and plans will be supported in the future.

How the hand-offs work

Google stresses that Orchestrator is not a router that guesses task difficulty from your first prompt and sends the whole job to one model. According to the models page, both models work inside the same task:

  • The planner plans. Each message starts with a read-only planning phase. The planner explores the codebase, resolves open questions and writes a plan that includes the conditions that should trigger re-planning. Questions and reviews get a direct answer from the planner.
  • Hydrogen does the work. Once the plan is ready, Hydrogen edits files, writes code, runs builds and tests, and checks results against the plan.
  • Control returns on friction. If a verification step fails twice, an interface behaves differently than expected, or the plan did not settle a design choice, Hydrogen hands control back. The planner repairs the plan and passes execution back.
  • The planner reviews. On longer tasks, and whenever you send a message while Hydrogen is working, Hydrogen reports progress and the planner decides whether to wrap up, continue or update the plan.

These hand-offs happen within a single response, with no manual switching.

How to select Orchestrator

In the Antigravity App:

  1. Click the model selector under the conversation prompt box.
  2. Choose Orchestrator (Carbon) or Orchestrator (Argon). Hover over the New badge to see which planner each option uses.
  3. Type your task. Orchestrator stays selected for the rest of the conversation until you pick another model.

In the Antigravity CLI, you have two methods:

  1. Run /model, select Orchestrator (Carbon) or Orchestrator (Argon), and press Enter. The choice persists across sessions.
  2. Pass the name in quotes at startup or in headless mode, for example agy --model "Orchestrator (Carbon)" or agy -p "Add input validation to the signup form and run the tests." --model "Orchestrator (Argon)".

Run agy models to list every model identifier available to your account. If you are new to the CLI, our Antigravity CLI 1.5 guide covers setup.

When to use Orchestrator or a single model

The models page calls Orchestrator the recommended default and gives this guidance:

Choice Best for Thinking level
Orchestrator Day-to-day feature work, bug fixes, refactoring and multi-file debugging Fixed
Gemini 4 Carbon or Argon Work where every step needs maximum reasoning, such as designing distributed systems, untangling race conditions or multi-repository migrations Selectable
Gemini 4 Hydrogen Package upgrades, deduplication, boilerplate, unit tests for well-defined functions, documentation updates and quick scripts Selectable: Low, Medium, High (default) or Extra (xhigh)

Orchestrator keeps a single agent in the conversation. For several agents working in parallel or checking each other, Google points to subagents, Swarm, Boost and Teamwork instead. Our guides to Antigravity /swarm and Antigravity Dynamic Workflows cover those multi-agent options.

For fixed roles in your own code, the launch post says you can wire Hydrogen and Carbon together with the Antigravity SDK.

Speed and quota

Google reports that Orchestrator reached near-Carbon task success rates on Terminal-Bench while using 24% less quota on average and finishing 19% faster. The Antigravity post cites Terminal-Bench 4.1, and the Google blog announcement cites Terminal-Bench 4.0. These are Google’s own figures.

Quota in Antigravity is not a fixed prompt count. The plans page says rate limits track the amount of work the agent does, so complex tasks use more. Google AI Ultra and Pro quotas refresh every five hours, subject to weekly limits, and other plans get a weekly quota. Pro and Ultra users can buy AI credits for overage, controlled by the AI Credit Overages setting (Never or Always).

Limitations

  • No effort control. /effort and --effort do not apply. Running /effort reports, for example, “Effort isn’t adjustable for Orchestrator (Argon).”
  • Gemini only. Both current options pair Gemini 4 models. Google says more models will be supported later.
  • Plan gated. Free and Google AI Plus users cannot select Orchestrator, and Carbon planning requires Google AI Ultra.
  • Mid-turn changes. If you switch models while the agent is running, it finishes the current turn with the previous model.

Frequently asked questions

What is Orchestrator in Google Antigravity?

Orchestrator is a model selector option in Google Antigravity that pairs a planner model, Gemini 4 Carbon or Gemini 4 Argon, with Gemini 4 Hydrogen. The planner plans and reviews, Hydrogen edits code and runs tests, and control passes back and forth within one response.

Which plans include Antigravity Orchestrator?

Google AI Ultra includes Orchestrator (Carbon) and Orchestrator (Argon). Google AI Pro includes Orchestrator (Argon) on paid plans or trials with eligible premium devices. The Free and Google AI Plus tiers do not include Orchestrator.

Is Orchestrator the same as a model router?

No. Google says Orchestrator does not classify your prompt once and send the whole task to one model. Both models work on the same task, and Hydrogen hands control back to the planner when it hits friction, such as a check that fails twice.

Can I set a thinking level for Orchestrator?

No. Orchestrator runs with a fixed configuration, so /effort and –effort do not apply. To choose a thinking level, select Gemini 4 Carbon, Gemini 4 Argon or Gemini 4 Hydrogen directly in the model selector.

How do I use Orchestrator in the Antigravity CLI?

Run /model and pick Orchestrator (Carbon) or Orchestrator (Argon), or pass the name in quotes at startup, for example agy –model "Orchestrator (Argon)". Run agy models to list the model names your account can use.

Does Orchestrator use less quota than Gemini 4 Carbon?

Google reports that in Terminal-Bench testing, Orchestrator reached near-Carbon task success while using 24% less quota on average and finishing 19% faster. Actual quota use varies, because Antigravity rate limits track the amount of work the agent does.

Sources

Related answers

Leave a Comment