Google shipped Orchestrator for Antigravity on November 12 — a feature that automatically routes work between Gemini 4 Carbon and Hydrogen within a single coding task. The headline numbers: near-Carbon quality, 24% less quota burned, 19% faster completion. If you’re on Google AI Ultra and using Antigravity as your primary coding environment, this directly changes your monthly cost structure.
Not a Router. A Handoff Loop. #
The important distinction is what Orchestrator is not. It’s not a classifier that reads your prompt and routes the whole request to one model. Both Carbon and Hydrogen work within the same task, passing control back and forth as the work demands.
Here’s what a typical session looks like. Carbon goes first — it explores the codebase in a read-only planning phase, resolves ambiguities, and writes a concrete implementation plan including the conditions that should trigger re-planning. If your message is a simple question requiring no code changes, Carbon answers directly and you’re done. For anything involving actual code, Carbon hands its plan to Hydrogen, which writes the implementation, runs tests, and iterates.
When tests fail or the implementation runs into something unexpected, control transfers back to Carbon for re-reasoning. The handoffs happen automatically inside a single response thread. You don’t change a setting or spin up a new session. It’s one conversation, two models, working in sequence.
The Numbers #
Google benchmarked Orchestrator on Terminal-Bench 4.1. The results: near-Carbon task success rates, 24% less quota consumed, 19% faster completion. For comparison, Orchestrator more than triples the benchmark score of Gemini 3.8 Flash (57% vs 18.2%), which gives useful context on how the underlying model capability stacks up.
The practical translation: if you’re regularly hitting Carbon quota ceilings on complex projects, Orchestrator extends your monthly runway without accepting a quality tradeoff. That’s the value proposition in one sentence.
| Metric | Carbon Only | Orchestrator |
|---|---|---|
| Task Success | Baseline | Near-Carbon |
| Quota Use | 100% | 76% (−24%) |
| Speed | Baseline | 19% faster |
| Terminal-Bench 4.1 | ~57% | ~57% |
The Pricing Lens #
Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens — 25% cheaper per token than Carbon’s current $2.00 / $10.00 pricing. That gap widens significantly on December 1, when Carbon’s introductory pricing expires. Hydrogen has no announced pricing change.
This matters for Orchestrator because Hydrogen handles the execution-heavy portion of tasks — writing code and running tests — which typically represents the bulk of output tokens. The quota savings compound with time, and compound further after Carbon’s price doubles next month.
There’s one catch worth flagging: Hydrogen’s thinking mode cannot be disabled. For certain task types, this means per-run costs can exceed what you’d spend running Carbon directly. Benchmark scores don’t surface this — you’ll want to watch your actual usage for the first few weeks.
Who Gets It #
Orchestrator with Carbon-as-planner is available to Google AI Ultra subscribers only. Pro subscribers get access to Hydrogen as a standalone model and an Argon-based orchestration option. Hydrogen is now Antigravity’s default model, replacing the previous default across both tiers.
The Ultra-only gate is a real friction point for adoption. For a feature positioned as a cost-saving measure, keeping it behind the highest-tier subscription creates an obvious irony. Developers on Pro plans pay less but can’t access the tool that reduces costs most aggressively.
The Operational Warning No One Is Talking About #
Hydrogen carries a “short-term GA” label, which means it’s subject to Google’s minimum 45-day retirement notice policy. That’s not a lot of runway if gemini-4-hydrogen is hardcoded into CI variables, infrastructure configs, or scripts across a team’s toolchain.
Google’s model naming has seen real churn — Flash became Hydrogen, naming conventions shifted, default models changed. If you’re integrating Antigravity programmatically or building CI pipelines around specific model endpoints, use alias-based references where the platform offers them. 45 days of migration notice for a production dependency is tight.
The Bigger Picture #
What Orchestrator represents matters beyond its quota numbers. Specialized models — one for reasoning, one for execution — are how production AI agent pipelines have worked for the past year. What Google did here is bring that architecture directly into the developer’s IDE, transparently, without asking developers to manage it themselves.
The pattern is straightforward: expensive frontier model for thinking, fast cheap model for doing. It’s the same logic behind microservices architecture — the right tool for the right subtask, coordinated automatically. Cursor and Copilot will ship their versions of this. Antigravity just went first.