Gemini 4 Hydrogen in Antigravity: Orchestrator Saves 24% of Quota

⚡ The gist in 5 seconds

• Google has enabled Gemini 4 Hydrogen in Antigravity and added an Orchestrator option to the model dropdown.
• Orchestrator hands the task off between Carbon and Hydrogen: the same results on Terminal-Bench 4.1, but with 24% less quota and 19% faster.
• Hydrogen is available to paid Google AI Pro and Ultra subscribers, Orchestrator (Carbon) — Ultra only.

🔍 What happened

On November 12, the Antigravity team published a post about two things at once: the Gemini 4 Hydrogen model and the Orchestrator mode. Hydrogen now sits in the model selector and works as a standalone model for code and agentic tasks. Orchestrator is a separate option in the same selector: it combines Hydrogen with the strongest models in the family, including Gemini 4 Carbon, for long-running tasks. The authors' numbers: 51 on the Artificial Analysis Intelligence Index (54 at max), 57% on Terminal-Bench 4.1 versus 18.2% for 3.8 Flash. Argon scored 50 in the same measurements.

Orchestrator's mechanics are not a static router that guesses complexity from the first prompt. Carbon starts in read-only mode: it parses dependencies, resolves ambiguities, and writes a plan with explicit replanning triggers. Then Hydrogen steps in — edits, package updates, tests. If it gets stuck, control returns to Carbon: inspect the state, fix the plan, continue. For those who want more automation, there's the Antigravity SDK: in the demo, Hydrogen agents triage incoming issues and write code, while Carbon gets large refactors, new features, and security checks. Hydrogen has four thinking levels: low, medium, high (the default), and the new xhigh.

🧠 Our take

Carbon and /swarm launched "earlier this week," and within a couple of days tens of thousands of swarms were run on Antigravity — so says the team itself. Our assessment: this news isn't about model quality, it's about the bill. Running a frontier model at every step of a routine task is expensive in quota, and Google openly calls it "burns frontier quota."

What this changes: the agent stops being "one model for the entire trajectory." Google is no longer selling a model, but a dispatcher — who plans, who writes, when to rebuild the plan. Minus 24% quota and plus 19% speed is an argument for those hitting subscription limits rather than benchmark scores.

Who wins. Ultra subscribers get Orchestrator (Carbon) right away, Pro gets only the (Argon) version. We believe the biggest winner is Google itself: dynamic task handoff within a single trajectory is a good reason to keep people on the higher-tier subscription instead of leaving for competitors.

⚠️ What we don't know

24% and 19% are vendor figures on a single benchmark, Terminal-Bench 4.1, with no published methodology and no number of runs. "Almost like Carbon" is a phrase from the blog; how "almost" is left unsaid. How many tasks break during the back-and-forth handoff — also unknown. The date in the post, November 12, 2026, looks odd: check the page yourself before citing it.

Prediction: within a quarter, Google will extend Orchestrator (Carbon) to paid Pro subscribers — the post explicitly mentions "additional models and plans."

Source: Google Antigravity Blog