Google's Antigravity Orchestrator is a new option for coding work that separates planning from execution inside a single session. It can move between Gemini 4 Carbon and Gemini 4 Hydrogen as the work changes, which means a task can use different models for different steps. Google launched Hydrogen on November 12, and access to the Orchestrator depends on your subscription. Here is what it does, who can use it and how to test it on a real task.
What Orchestrator Does
The core idea is a split between thinking and doing. Carbon explores the codebase, identifies dependencies and produces a plan in a read-only phase. Hydrogen then takes over to edit code, update packages and run tests. If execution runs into friction, control can return to Carbon, which inspects the current state and repairs the plan.
That is different from a static router that picks one model for an entire job. Google says Orchestrator can change models as the work changes. For parallel work, Carbon coordinates while Hydrogen powers worker and verifier agents, each working in an isolated workspace. The design aims to keep the expensive planning step separate from the many smaller edits that follow.
Who Can Use It
Access depends on the product and the plan:
- Gemini app, free users: Hydrogen is applied automatically when Google judges it suitable for a prompt.
- Paid Google AI Pro and Ultra subscribers in Antigravity: Hydrogen is rolling out to these plans.
- Orchestrator options: Ultra subscribers can use the Carbon and Argon versions, while Pro subscribers get the Argon version.
- Enterprise customers: Both models are available through Gemini for business.
- Developers: Hydrogen is available through the Gemini API and Google AI Studio, and Android Studio access is on the paid tier.
The pages reviewed for this guide list no prices, so check your current plan before planning a workflow around Orchestrator.
A Practical Test Plan
Start with a task that has a clear finish line. A bug with a failing test or a package upgrade with a known test suite works well, because the outcome is easy to check. Orchestrator's read-only planning phase suits multi-file work where the first risk is misunderstanding the codebase.
Next, review the plan before execution begins. The planning step is the cheapest point to catch a wrong assumption, such as a misread dependency or an unnecessary file change. If the plan is wrong, correct it there rather than after several rounds of edits.
During execution, watch for model switches. Confirm which model handled which step, and if a step fails, check whether control returned to Carbon for plan repair. Those transitions are where you can learn the most about how the system behaves on your code.
Finally, review the diff and test results yourself. Agents that run tests can still write or modify tests in ways that pass for the wrong reason, so read the changes, not just the green checkmark.
Measuring Whether It Helps
Google's Antigravity team reports that Orchestrator achieved task success close to Carbon's on Terminal-Bench 4.1, using 24% less quota and finishing tasks 19% faster on average. These are Google's own figures, and independent results may differ by codebase and task type. The useful measure is your own: run the same task with and without Orchestrator, and compare the time, quota use and the number of corrections you had to make.
Quota savings only matter if quality holds. A cheaper run that needs two extra review cycles is not a win. Track the total time from request to merged change, not just the agent's runtime.
Setting Up a Repeatable Check
Once you have one good run, turn the process into a checklist you can repeat. Write down the task, the files you expect to change, the tests that must pass and the review steps before merge. Keep a short log of each run's model transitions, quota use and corrections. After a few tasks you will know whether Orchestrator helps with your type of work, such as refactors, dependency updates or test writing, and where it adds overhead.
Common Pitfalls
- Vague scope. A broad request gives the planner too much room. Name the files, tests or behavior you expect to change.
- Parallel work without merge review. Isolated worker workspaces are useful, but each one needs careful integration review.
- Trusting the plan blindly. A read-only plan can still be wrong about how the code behaves at runtime.
- Ignoring plan tier limits. If your subscription does not include the Orchestrator option you expected, the workflow will not be the one described here.
Why It Matters
Routing models within a single task is a shift from choosing one model for a whole job. It gives developers a way to put strong planning where it matters and cheaper execution where it does not. Users who review plans, watch model transitions and measure results against their own work will get the most value from it. For everyone else, the feature is a reminder that the best model for a task may change partway through it.
