Gemini 4 Hydrogen and Orchestrator

Orchestrator is an option in the model selector of Antigravity, Google's agentic coding tool, that runs two models inside each turn instead of making you pick one. A larger planner, Gemini 4 Carbon or Gemini 4 Argon, plans and reviews, while the fast, lightweight Gemini 4 Hydrogen does the implementation. According to the Orchestrator docs, every message starts with a read-only phase in which the planner explores the codebase, resolves ambiguities and writes a plan that names the conditions that should trigger re-planning. Hydrogen then edits files, runs builds and tests, and checks the results against that plan. When it hits friction, such as a verification step that fails twice or an interface that behaves differently than expected, it hands control back to the planner, which repairs the plan and returns execution to Hydrogen. These hand-offs happen automatically within a single response.

The reason to choose it is where the expensive reasoning gets spent. Google's Antigravity announcement argues that running a frontier model like Gemini 4 Carbon on every step of routine work slows the loop and burns frontier quota. It contrasts Orchestrator with a static model router, which guesses difficulty from the first prompt and sends the whole task to one model. A long-running agent often discovers midway that a step is harder or easier than it looked, and a router cannot revise its choice, while Orchestrator can pass control in either direction. The docs recommend Orchestrator as the default for feature work, bug fixes, refactoring and multi-file debugging. They recommend picking Carbon or Argon directly when every intermediate step needs maximum reasoning depth, and Hydrogen alone for jobs like package upgrades or boilerplate that need no separate planning phase.

The evidence that the split pays off is a single result Google reported about its own product. The Antigravity post says that on Terminal-Bench 4.1, Orchestrator reached "near-Carbon task success rates" while on average using 24% less quota and finishing 19% faster. Google gives no absolute success rates, so a reader cannot tell how large the gap behind "near" is, and the post does not state the baseline outright, though the docs describe the savings as relative to running the planner alone. Google's main Hydrogen launch post attributes the same figures to Terminal-Bench 4.0, so even the benchmark version is unsettled between Google's two posts. Until an independent run publishes success rates, the efficiency gain is a plausible vendor claim rather than an established one.

Two practical limits also bear on the choice. Orchestrator runs with a fixed configuration, so you cannot set a thinking level for it; selecting Carbon, Argon or Hydrogen directly restores that control, including Hydrogen's new Extra (xhigh) level. Access depends on your plan: Google AI Ultra subscribers get both Orchestrator (Carbon) and Orchestrator (Argon), while paid Google AI Pro subscribers, and Pro trials on eligible premium devices, get only Orchestrator (Argon).