Running the most capable model on every step of a coding task is slow and eats through your allowance, but routing the whole job to a cheaper model risks it stalling halfway. Google's answer in Antigravity is Orchestrator, a new option in the model menu that hands work back and forth between Gemini 4 Carbon and the new, faster Gemini 4 Hydrogen within a single task. If you pay for Antigravity, it is the setting most likely to change how far your quota goes.
The idea, as Google's Antigravity blog describes it, is that a task's difficulty changes as it unfolds. A common approach elsewhere is a static router that judges difficulty from the first prompt and sends everything to the cheapest model that should cope. Orchestrator instead switches in both directions. Carbon might start by reading the codebase, resolving ambiguities and writing a plan with conditions for re-planning; Hydrogen then edits files, updates packages and runs tests; if Hydrogen runs into trouble, Carbon takes over again to inspect the state and repair the plan.
Google's numbers are specific. On Terminal-Bench 4.1, Orchestrator reached close to Carbon's task success rate while using 24% less quota on average and finishing 19% faster. Hydrogen on its own scores 51 on the Artificial Analysis Intelligence Index, above Argon's 50 and 54 at its highest setting, and 57% on Terminal-Bench 4.1, against 18.2% for Gemini 3.8 Flash. Hydrogen offers four thinking levels: low, medium, high by default, and a new xhigh. For finer control, developers can wire the two models into separate roles with the Antigravity SDK.
The approach mirrors what many developers already do by hand in Claude Code or Cursor: plan with a strong model, execute with a cheap one. Building the switch into the tool removes that chore, and doing it mid-task rather than at the start is the sensible part. Availability is tiered: Hydrogen is rolling out to paid Google AI Pro and Ultra subscribers, Orchestrator (Carbon) is for Ultra only, and Pro subscribers get Orchestrator (Argon).
The caveats are the usual ones for vendor launches. Near-Carbon success is Google's own measurement on one benchmark, and near is not defined in the post. Google does not say how often the hand-backs happen or how the system decides to make them, which matters if you need to predict cost. Independent results on real projects will be the better guide.