Image: Google
Why it mattersA team can keep Carbon for planning. Hydrogen handles the routine work. Nobody has to switch models by hand.
Routine coding tasks still run on the frontier model when a developer picks one. That costs more quota than a routine task needs. Google Antigravity now adds a second, faster model. A controller decides which of the two runs each step.
Google added Gemini 4 Hydrogen to Antigravity on 12 November 2026. It also introduced an Orchestrator option in the model picker. Orchestrator sends planning and analysis to Gemini 4 Carbon, the frontier model. The smaller steps go to Gemini 4 Hydrogen. When Hydrogen gets stuck, the work escalates back to Carbon.
What Google reports
Google says Orchestrator reaches near-Carbon task success rates. It uses 24 percent less quota. It finishes tasks 19 percent faster on average. Google gives these numbers as its own, measured on Antigravity. Hydrogen on its own scores 57 percent on Terminal-Bench 4.1. No comparison against any third-party agent setup is reported.
Where it fits
Hydrogen is listed for routine work. Google names dependency upgrades, refactors across shared utilities, and alt text on image assets. Carbon stays the choice for long planning passes. One careful step there avoids many wrong ones later.
The Orchestrator option is new. A team used to pick one model for a whole session. Now a developer can hand that choice over for a single task.
The model picker in Antigravity's app, IDE, CLI, SDK and extensions carries both names. A plan still shows which steps ran on which model. So the quota line in a finished task breaks down by model, rather than hiding both in one total.
What a team should check
The 24 percent figure is a measurement Google reports on its own product. No external comparison is included. A buyer running an evaluation would want the same tasks on Carbon alone, on Hydrogen alone, and on Orchestrator. The three totals can then be compared under one workload. Reading between Google's own published runs is not the same thing.
Terminal-Bench 4.1 is the benchmark Google names. Hydrogen's 57 percent should be read as the number at publication, not a durable ranking.
Orchestrator will also make a difference only when a task has room to split. A short, well-scoped change may stay on one model from start to end. The saving shows up in a long session that mixes a few careful decisions with many routine edits. That is the shape where a frontier model is being paid for every step. The point of the pair is that Carbon is free to handle only the steps that need it.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
