On 11/12/2026 Google added two names to Antigravity: the Gemini 4 Hydrogen model and a selector called the Orchestrator. The second one is what matters. Instead of locking in one model and riding it to the end of the task, the Orchestrator swaps models inside the same task: a larger model plans and reviews, Hydrogen edits, runs packages and tests, and when friction shows up the work goes back to the larger model.
The promise comes with a number. According to Google itself, the Orchestrator gets close to the large model's result while using, on average, 24% less quota and finishing 19% faster.
That number says more than it seems. A router that swaps models to save quota only exists because quota is scarce. So the useful question isn't "is the Orchestrator good?", it's "why does the tool need to ration the model in your place?".
Below are the two sides of the launch (what Google claims and what devs pushed back with on Hacker News), the technical part of Hydrogen, and how to make that same choice by hand in Verboo Code.
What does the Antigravity Orchestrator do?
It swaps models mid-task, without you asking. In Google's description, Carbon (the larger model) handles reading, planning and re-planning, and Hydrogen handles edits, package updates and tests. When the task hits friction, it goes back to Carbon.
The difference from a typical router is when the decision happens. A static router guesses difficulty before starting. The Orchestrator decides during execution, step by step.
Gemini 4 Hydrogen's numbers
Hydrogen is the fast half of the pair. It ships with four reasoning levels: low, medium, high (default) and a new xhigh. The numbers below are Google's claims on the launch page.
| What Google claims | Number |
|---|---|
| Intelligence Index (Artificial Analysis) | 51 (54 at the highest setting) |
| Terminal-Bench 4.1 | 57%, against 18.2% for Gemini 3.8 Flash |
| Orchestrator quota saving | 24% less, with results close to Carbon |
| Orchestrator time | 19% faster |
One caution before repeating those numbers: on the Artificial Analysis index checked on 11/13/2026, Hydrogen shows up only as a fast tier, around 1.3 minutes per task, and with no published index score. The 51 is Google's claim, not a score Artificial Analysis displays today. Cite it as "Google says", never as an independent measurement.
Where is the Orchestrator available?
Hydrogen lands on the paid Google AI Pro and Ultra plans. The Orchestrator is tighter: "Orchestrator (Carbon)" and "Orchestrator (Argon)" only on Google AI Ultra, and on Pro only "Orchestrator (Argon)". For companies, both come through Gemini for business.
What did devs point out at launch?
The reaction on Hacker News (445 points and 133 comments on the launch thread) praised one thing and criticized three.
The praise is direct. Hydrogen's speed, and not having to alternate model and effort by hand at every step. One comment sums up the mode as "I don't have to keep switching between models and efforts".
The first criticism is the naming. On the periodic table the order is Hydrogen, then Carbon, then Argon. In the size of Google's models, the real order is Hydrogen, then Argon, then Carbon. That confuses even people who follow model launches closely.
The second is what was left out of the entry plan. On paid Pro you get Hydrogen, but the Orchestrator with Carbon only on Ultra. One comment sums up the entry tier as "a paid demo".
The third is lock-in. The Google subscription doesn't let you take the model into your own harness, which makes the product less useful for anyone whose flow is already built. And much of the thread argues this routing should be the default behavior, not a menu item you pick.
Should you let the tool pick the model for you?
If your bottleneck is quota, yes. A router that uses the large model only where it's needed beats you remembering to switch to the cheap model on easy steps. There, the Orchestrator fixes a real annoyance, and the 24% quota saving Google reports is exactly the proof that the problem exists.
But the saving is the point. The Orchestrator wasn't designed to make the task smarter, it was designed to stretch it inside a ceiling. If you program without that ceiling, you don't need a router. You need the best model to stay on the task, when the task calls for it.
That's the design difference between the two tools. The Orchestrator decides for you so it fits the quota. In Verboo Code the choice is yours, and there's no quota to fit: /model picks the model, /effort picks how much it thinks, and tokens are unlimited on the plan.
| Attribute | Orchestrator (Antigravity) | Verboo Code |
|---|---|---|
| Who picks the model | The tool, during the task | You, with /model |
| Reasoning level | low, medium, high, xhigh, set by the model | low, medium, high, max, with /effort |
| Why it swaps models | To save quota | No quota to save |
| Interface | Antigravity IDE | CLI, in the terminal |
| Entry plan | Google AI Pro (paid), Orchestrator with Carbon only on Ultra | Junior, R$ 99/month |
| Usage limit | Quota per Google plan | Unlimited tokens on the plan |
How do you pick the model in Verboo Code?
In Verboo Code the choice is a command, not a router. Inside the session, just type:
/model opus
/effort high/model opens the model picker when you run it with no argument, and takes direct aliases like opus, sonnet, haiku, best and opusplan. /effort sets how much the model thinks, from low to max, and takes auto to let the model decide the level. When the task needs more speed, /fast turns on fast mode.
The underlying difference is this: here you switch models when you want to, not when the quota allows.
The Orchestrator solves quota by swapping models internally. If quota isn't your problem, Verboo Code gives you unlimited tokens on the plan, so the good model stays on the task from start to finish.
