Antigravity: Google Orchestrates Gemini with Hydrogen

Google has activated Orchestrator in Antigravity to manage multiple Gemini models within a single session, with Gemini 4 Hydrogen running. The company specifies subscription access and pricing for Hydrogen, while analysts see potential for productivity gains and token savings, accompanied by increased requirements for traceability and security.

Access, Subscriptions, and Pricing Announced by Google

Orchestrator is accessible via the model selector located under the conversation input area of Antigravity. Google currently offers two configurations: with Carbon as the planner for Google AI Ultra subscribers, and with Argon as the planner for Google AI Pro and Ultra subscribers. The selected configuration remains active throughout the conversation until another model is chosen. Developers can also select Gemini 4 Hydrogen alone for tasks that do not require planning.
Gemini 4 Hydrogen offers a context window of 1 million tokens. Its pricing is $1.50 per million input tokens and $7.50 per million output tokens.

A Multi-Model Flow with Automatic Control Transfers

Orchestrator combines a planner, such as Gemini 4 Carbon or Argon, with Gemini 4 Hydrogen, which Google describes as a lightweight execution model. The planners navigate the codebase, identify dependencies, and design a plan, then delegate the execution of tasks such as file modifications, code generation, and builds and tests to Hydrogen. Unlike traditional routing, which assigns the entire task to a single model, Orchestrator repeatedly transfers control between planning and execution phases as progress is made.
The planner resolves ambiguities before handing control back to Hydrogen. In case of difficulties, such as repeated test failures or unexpected behavior, Hydrogen returns control to the planner, who can revise the plan. These back-and-forth exchanges occur automatically within a single response, without manual intervention to switch models.

Productivity and Costs: Gains Anticipated by Analysts

The ability to orchestrate multiple models within a single session, with task-type distribution, is presented as a productivity lever. Pareekh Jain anticipates an increase in coding speed, improved debugging and quality, as well as better completion of complex tasks, comparing the setup to a senior architect and a very fast coder duo. Amit Kumar Jena notes that developers often abandon tasks when agents encounter complex requirements or repeated failures.
Regarding costs, Jena highlights a potential for token efficiency: relying on a single high-performing model for everything would incur high costs for routine tasks, while a lighter model would penalize complex tasks. Automatic assignment based on the required reasoning could reduce unnecessary model usage and make costs more predictable. For CIOs overseeing large teams, Jain believes this can improve expense management and help expand assisted development without quickly exhausting usage caps or increasing bills.

Traceability and Security: Requirements Raised by Experts

Pareekh Jain warns about accountability and auditing when a flow involves multiple models: CIOs need to know which model made a decision or modified a segment of code. Without model-level logs and fine-grained attribution, auditing generated code, investigating errors, establishing accountability, or understanding model selection becomes difficult. Amit Kumar Jena notes that at this stage, Antigravity's documentation does not describe step-by-step logs identifying the model responsible for each action and recommends verifying the traceability of changes back to the model and plan followed.
On security, Jena reminds that each model brings its own behaviors and risks regarding access to source code, secrets, and production credentials. He advocates for assigning dedicated identities to agents, restricting credentials to relevant tasks, and requiring human approval before any code merges. He also encourages examining the management of intellectual property by each model: where prompts and code are processed, data retention, and contractual coverage for all models that Orchestrator may invoke.