On Thursday, Google introduced Gemini 4 Hydrogen alongside Orchestrator, a fresh capability within its Antigravity development environment that fundamentally alters how coding agents deploy AI models. Instead of relying on a single high-end model throughout an entire session, the system intelligently alternates between Hydrogen and the more advanced Gemini 4 Carbon as work progresses.

According to Google's testing, this hybrid approach delivers performance nearly equivalent to running Carbon independently while reducing quota consumption by 24% and accelerating task completion by 19%.

How Orchestrator handles the work

The model-switching mechanism operates through Orchestrator, which assigns Carbon to tackle computationally intensive reasoning tasks such as system architecture and debugging. Hydrogen takes responsibility for implementing code modifications and executing tests.

In practice, Carbon might analyze a repository to determine how to add a feature, then transfer execution to Hydrogen for the actual implementation. When test failures emerge, Carbon resumes control to reassess the strategy before handing work back to Hydrogen.

Clement Farabet, VP of AI engineering at Google DeepMind, described the pairing as "a huge unlock for agentic coding" in a post on X. Varun Mohan, heading Google's Antigravity team, noted in a separate post that combining Hydrogen with Orchestrator and swarms has gained significant traction "very popular internally."

Where Google's numbers don't match

Inconsistencies emerge when comparing Google's benchmark disclosures. The primary Gemini announcement references Terminal-Bench 4.0, while the Antigravity announcement cites version 4.1.

Additionally, Google has not released Carbon's individual success rate, provided per-task performance breakdowns, or shared sufficient information about how Orchestrator makes routing decisions. This lack of transparency makes it difficult to pinpoint the sources of the claimed efficiency improvements.

…hasn't published Carbon's standalone success rate… or enough detail about Orchestrator's routing decisions to show where the reported efficiency gains actually come from.

Hydrogen outscores a pricier model

Google positions Hydrogen as a "daily driver for agentic workflows," priced at $1.50 per million input tokens and $7.50 per million output tokens. These rates apply uniformly across its complete million-token context window.

The model is accessible via Google AI Studio, the Gemini API, the Gemini app, and Google Search.

On the Artificial Analysis Intelligence Index, Hydrogen achieves a score of 51 at its standard high thinking level, narrowly surpassing the costlier Gemini 4 Argon at 50. Increasing the reasoning intensity to Google's new xhigh setting lifts Hydrogen's score to 54, though this enhancement can raise token usage and computation duration.

Argon held the position of Google's leading model when it rolled out on September 30, but Gemini 4 Carbon outperformed it across most public benchmarks following its October 28 debut. Hydrogen's single-point advantage may seem marginal, yet it underscores the rapid advancement of Google's budget-friendly models in matching capabilities that represented cutting-edge performance mere weeks earlier.

Google further reports that Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash under default conditions, achieving a Terminal-Bench 4.1 score of 57% compared to Flash's 18.2%. The company also claims more than a fivefold decrease in hallucinated content.

The cost of switching models

Google has not revealed the criteria Orchestrator uses to determine when to switch between models or the token and latency expenses associated with these transitions.

Currently, Orchestrator access is restricted to Google AI Ultra subscribers, with expansion to additional models and subscription levels in the pipeline. Hydrogen is rolling out to paying Antigravity users, with full public availability anticipated.

Google hasn't disclosed how Orchestrator decides when to switch models or what those handoffs cost in additional tokens and latency.