DEV Community

Cover image for The Week Gemini 4 Carbon and Hydrogen Landed: Antigravity's Orchestrator and /swarm, Explained
SeaOtter ( Kentaro Mori )
SeaOtter ( Kentaro Mori )

Posted on

The Week Gemini 4 Carbon and Hydrogen Landed: Antigravity's Orchestrator and /swarm, Explained

1. Introduction

When you hand heavy work to an AI agent, you keep running into the same trade-off: the smart models are slow and burn through your usage quota, while the fast models get stuck exactly where it matters.

Between November 10 and 13, 2026, Google shipped its answer to that problem all at once: Gemini 4 Carbon, its strongest model; Gemini 4 Hydrogen, a fast new model; and two new features in its development environment Antigravity — /swarm and Orchestrator.

Since Gemini 3.8 Flash arrived in September, Google's model lineup had felt like it was standing still. This week's releases refresh it completely. This article summarizes what changed and how the pieces work, based only on Google's official blog posts.

2. What was released

The six main changes this week:

  • Gemini 4 Carbon (November 10) — Google's strongest model, built for long-horizon agentic work, complex coding and debugging, and scientific reasoning. In subscriptions it is available only on Google AI Ultra; the API is open to paying developers.
  • Gemini 4 Hydrogen (November 12) — The fast successor to Gemini 3.8 Flash, built for everyday coding and agent execution, with four thinking levels. In Google Search it is available to all users.
  • /swarm (November 10, preview) — An Antigravity feature that turns a single prompt into a coordinated team of agents whose progress you can track in real time.
  • Orchestrator (November 12) — A new option in Antigravity's model selector that dynamically switches between models within the same session. At launch it was for Ultra subscribers, pairing Carbon as the planner with Hydrogen for execution. On November 13 it also reached Pro subscribers, where it pairs Argon as the planner with Hydrogen. Orchestrator with Carbon remains Ultra-only.
  • Dynamic Workflows (November 13, /run-workflow) — Lets the agent generate and run a short Python script that coordinates many agents, running phases in parallel with parallel() or in sequence with pipeline().
  • Smaller Antigravity updates — A font size setting, split-view plugins, approval before the browser agent downloads a file, and a layout tailored to your kind of work. On November 11, Carbon also became selectable when signing in with Google Cloud credentials.

3. How it works

Both /swarm and Orchestrator are built on the same idea: don't make one model do everything.

In /swarm, Carbon takes the roles that demand deep reasoning — architectural design, coordination, and auditing. Faster, lighter models act as workers and verifiers in parallel, each in an isolated workspace, and the verifiers test every change. According to Google, tens of thousands of swarms were launched in the first few days.

A /swarm session in Antigravity: agents with split roles analyzing a repository in parallel (author's screenshot)

Orchestrator divides roles inside a single session. Carbon starts in a read-only phase: it explores dependencies, resolves ambiguities, and builds a plan with explicit re-planning triggers. Hydrogen then takes over to write edits and run tests. If Hydrogen hits friction, it hands back to Carbon, which inspects the state, repairs the plan, and lets Hydrogen resume.

Unlike a static router that picks one model up front, Orchestrator moves back and forth as the work evolves. In Google's measurements on Terminal-Bench, Orchestrator reached near-Carbon task success rates while consuming 24% less quota on average and finishing 19% faster.

Orchestrator launched for Ultra subscribers, but on November 13 it also reached Pro subscribers: on Pro, Argon plans and Hydrogen executes.

How Orchestrator works

Dynamic Workflows (/run-workflow) also arrived the same week, on November 13. When you run /run-workflow, the agent writes a short Python script that coordinates many agents. The script runs after your approval, moving through phases in parallel with parallel() or in sequence with pipeline().

Each agent works in its own conversation and returns a structured result matching a JSON Schema, so your main conversation's context stays clean. You can set a model and an isolated workspace per agent, and save the script to a file to run it again.

Orchestrator plans and routes inside a single conversation. Dynamic Workflows, by contrast, controls the steps with a script — a predictable approach.

/swarm has agents coordinating with each other on a shared message board. Dynamic Workflows fits work like applying the same check across a list of files or packages.

4. My take

The /swarm idea itself is not new. But with Gemini 3.8 Flash as the worker, the team felt underpowered — there was not much to gain from coordinating weak agents.

That changes now that Hydrogen can do the execution while Carbon or Argon sits in the coordinator's seat. A fast, capable executor paired with a deep-thinking lead makes orchestration practical, and in my view it is a strong lineup against Anthropic and OpenAI.

After months of standing still since 3.8 Flash in September, Google's models feel refreshed — and, for the first time in a while, on par with the competition.

5. Caveats

The caveat is consumption. Running several models in parallel structurally costs more tokens and quota. Carbon itself is designed to use more tokens than Argon to explore thoroughly and run extended verification loops, as Google's blog notes.

Orchestrator's 24% saving is a mechanism for keeping that cost down. Where you spend the frontier model, and which work you leave to the lighter one, will matter more in day-to-day development.

Also note that /swarm is still a preview, and Carbon's API price is an introductory one that may change.

6. Summary

  • This week Google shipped Carbon and Hydrogen, completing the Gemini 4 lineup.
  • Antigravity gained /swarm and Orchestrator, making multi-model setups — a coordinator plus executors — a standard option.
  • The open problem is consumption: using each model where it counts takes deliberate design.

The same week also brought pre-orders for Gemini-powered intelligent eyewear. The collections, made with Gentle Monster, Warby Parker, and Samsung, go on sale December 10. Google also announced Guided Vision, a conversational audio description feature.

The same week, on November 11, also brought the Gemini Decisions API, a decision-only endpoint powered by Gemini Decision 1 that generates no text. One multimodal input of up to 256K tokens can be scored against up to 1,000 yes/no, choice, or rubric-scoring questions in a single pass, returning probabilities and confidence scores. Latency is 150–250 ms end to end, pricing is $0.10 per 1M input tokens with no output tokens billed, and a free tier is available in AI Studio — built for tool selection, policy checks, and model routing, where only the decision is needed.

Also on November 10, DeepMind announced Gemini Robotics 2.5, its most capable vision-language-action model, evaluated across more than 200 tasks. In a real-world proof of concept, Apptronik's Apollo 2 humanoid performed automotive kitting on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, with a 96% success rate (a proof-of-concept figure from Google). Development was accelerated using Gemini 4 Argon, and the model is now rolling out to robotics partners.

References

Top comments (0)