Google's Antigravity Orchestrator Adds Gemini 4 Hydrogen

By Amara Osei

Google's new Antigravity Orchestrator splits one coding session across several Gemini models, with Gemini 4 Hydrogen priced at $1.50 per million input tokens and $7.50 per million output tokens.

Antigravity, Google's AI-powered development environment, now coordinates planning, coding and review work automatically. The release places the lighter execution model, Gemini 4 Hydrogen, behind planners such as Gemini 4 Carbon or Argon. Carbon and Argon explore the codebase and set the approach; Hydrogen edits files, writes code, and runs builds and tests.

How does Orchestrator split the work?

Conventional model routing picks one model from the prompt and gives it the entire job. Orchestrator changes that sequence. The planner first scans the codebase, maps dependencies, resolves ambiguities, and produces an implementation plan before handing control to Hydrogen.

When execution runs into trouble, such as repeated test failures or unexpected behavior, Hydrogen returns control to the planner. The planner reassesses the execution and revises the plan, then Hydrogen resumes implementation. Google said these handoffs happen automatically inside a single response, so developers neither switch models manually nor configure the transitions.

In plain terms, the planner decides and Hydrogen executes, with a feedback loop whenever the code fails.

Who can use Orchestrator, and what does it cost?

Developers reach Orchestrator through the model selector below Antigravity's conversation prompt box. Google AI Ultra subscribers can pair Carbon as the planner, while Google AI Pro and Ultra subscribers can pair Argon. The selected configuration stays active throughout the conversation until developers choose another model.

Developers can also select Gemini 4 Hydrogen on its own for tasks that need no planner. The model carries a 1 million-token context window. Google charges $1.50 per million input tokens and $7.50 per million output tokens, so output costs five times the input rate.

The commercial choice comes down to subscription tier and whether a task needs a planner at all.

Can enterprises spend less on coding?

Pareekh Jain, principal analyst at Pareekh Consulting, expects the multi-model setup to improve delivery. "This is like pairing a senior architect with a superfast coder. This could improve coding speed, debugging, code quality, and completion of complex tasks," he said.

He also points to a familiar failure mode. Amit Kumar Jena, AI development head at Kanerika, said developers often abandon tasks when AI coding agents struggle with complex requirements, repeated test failures, or unexpected code behavior.

Jena sees token efficiency as the larger prize for developers and CIOs. "By automatically assigning work to models based on the reasoning required, Orchestrator could help reduce unnecessary model usage and make AI coding costs more predictable," Jena said.

He warns that one capable model on every task risks overspending on routine work, while a lightweight model everywhere can weaken performance on complex tasks.

For CIOs overseeing large engineering teams, Jain expects tighter control over AI spending and room to scale AI-assisted development without exhausting usage limits or driving up costs.

The cost case rests on matching each task to the cheapest model that can still handle it.

What governance gaps remain?

Jain warned that coordinating several models inside one development workflow complicates accountability, auditing and security. He said CIOs need visibility into which model made a particular decision or code change.

Without model-level logs and clear attribution, he said, enterprises will struggle to audit AI-generated code, investigate errors, establish accountability after failures, or explain why a model was selected for a task.

Jena said Antigravity's documentation does not currently describe per-step logs identifying which model performed each action. He recommends that enterprises test independently whether they can trace each code change to the model that produced it and to the implementation plan it followed.

Each model also introduces its own behavior and risks when it accesses source code, secrets and production credentials. Jena recommends giving coding agents dedicated identities, restricting credentials to the specific tasks they perform, and requiring human approval before code changes merge.

He further advises checking how every model handles intellectual property: where prompts and source code get processed, what data providers retain, and whether enterprise agreements cover every model Orchestrator can invoke.

Without model-level audit trails and narrow credentials, enterprises cannot fully explain or contain what their agents did.

Google has made multi-model coding a managed workflow, but enterprises still need to prove they can trace it.