# Gemini 4 Carbon and /swarm turn one prompt into a team

**Summary:** Gemini 4 Carbon opens to paying users alongside Antigravity's /swarm preview. Google's 3D demos show the appeal of coordinated agents, with accuracy and total project cost still to check.

- Canonical: https://markhuang.ai/news/gemini-4-carbon-swarm-agent-team
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-10
- Section: News
- Tags: Gemini 4 Carbon, Google Antigravity, Multi-Agent Systems, Agentic Coding, 3D Reconstruction
- Source: [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![Three mechanical drafting arms assemble an Eiffel Tower model above architectural drawing sheets](https://cdn.markhuang.ai/news/gemini-4-carbon-swarm-agent-team/hero.webp)

*Conceptual illustration of agents building from plans. Google reports an Eiffel Tower construction exhibition made from historical photos and engineering plates; this artwork is not an image of that demonstration.*

Give an AI team seven planning PDFs and ask for a building you can explore. In [Google's November 10 announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/), a Gemini 4 Carbon team used 206 extracted engineering drawing views to build an interactive 3D reconstruction of its Platform 37 office. Google says the project began with one prompt and ran across 10 overnight loops. It is a company demonstration, but the task is easy to understand: turn a pile of documents into something a person can inspect spatially.

Carbon is now available through the API in Google AI Studio and Gemini for business for paying developers and enterprises. Google is also rolling it out to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. Alongside it, Google is previewing `/swarm` in Antigravity's App and CLI. That feature turns a prompt into a coordinated team of agents that can divide the work and check changes.

The appeal is being able to hand over a project that takes repeated attempts. A team can read the source material, build from it, and check its work as it goes. Carbon and `/swarm` give users a way to delegate that longer process while watching how the work develops.

## What the team adds to the model

Carbon is the model. `/swarm` is the system that organizes agents around a project. The distinction matters because you can assign different models to different roles; using Carbon does not require every worker to run it.

The [Antigravity team's announcement](https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity/) describes a coordinator that creates specialist workers and separate verifiers. Workers edit in isolated Git worktrees, separate working copies of a software project, so their simultaneous edits do not collide. Verifiers run tests and challenge proposed changes. Everyone can communicate through a shared message board that keeps earlier messages.

```mermaid

flowchart TD
    P[User prompt] --> C[Coordinator]
    C --> W[Specialist workers in separate worktrees]
    C --> V[Verifiers running tests and reviewing changes]
    C <--> B[Shared message board]
    W <--> B
    V <--> B
```

The interface lets users inspect agents' activity, steer individual agents, and replay the team's progress. That makes a long run easier to investigate: you can look at how the team reached a result as well as the final handoff. Those records would be useful when a polished output contains a questionable assumption.

## Documents become something you can explore

[Google's Eiffel Tower example](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) starts with historical photos and engineering plates. The reported result is an interactive exhibition of its construction. Together with the office reconstruction, it shows why this kind of agent project is worth trying. You can explore a structure instead of piecing it together in your head from separate drawings.

An educator could use that kind of exhibition to explain how a structure took shape. Someone reviewing a building could explore the model, then return to the original drawings to check a detail. These are reasons to try the approach, rather than evidence that the demonstrations are accurate enough for professional design decisions.

Google's announcement does not give the office model's reconstruction accuracy or total token bill. It reports 10 overnight loops, so one prompt should not be read as one instant result. The useful promise is less manual coordination across a project that takes time. Someone still needs to check the model against the drawings.

## Put the extra reasoning where it earns its cost

Google says Carbon uses more tokens than Argon to explore problems and run extended verification. It still recommends Argon for most software engineering and knowledge work. Antigravity's model selection lets users put Carbon on architecture, coordination, or difficult reviews while assigning routine implementation to faster, lighter models.

There is a concrete tradeoff in [CursorBench 4.0's published results](https://cursor.com/cursorbench), which evaluate agents on ambiguous tasks involving multiple files from real Cursor sessions. Carbon at High scores 60.4%, versus Argon at High at 52.9%. The listed average cost per task is $9.22 for Carbon and $4.57 for Argon. The stronger result comes with a higher bill in that evaluation. Cursor also cautions that small score differences may not be statistically meaningful.

Other evaluations ask different questions. [Artificial Analysis's October 28 assessment](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1) put Carbon first on its Intelligence Index v5 but second on its Coding Agent Index, behind Claude Opus 5.5. That broader lead does not settle which model to use for a particular coding job.

For a pilot, I would choose a project whose completion I can check against the source material or existing tests. The comparison would include the whole team's spend and the effort needed to review its output. Antigravity explicitly warns that parallel agent teams consume quota much faster than a standard chat.

## Access has widened, and the discount has a deadline

Google lists introductory public API prices of $2 per million input tokens and $10 per million output tokens. Its footnote now says that period expires on November 30, 2026. Afterward, the public Carbon model costs $4 per million input tokens and $20 per million output tokens. Those are token prices, not subscription prices. A swarm's bill depends on how much its agents do.

That updates two open questions in my [earlier Carbon versus Argon analysis](https://markhuang.ai/news/gemini-4-carbon-argon-upgrade): broader access has arrived, and the discount now has a stated end date. Access still depends on the product and plan. In the [launch discussion on Hacker News](https://news.ycombinator.com/item?id=50442303), commenters report different model menus and ask about subscription eligibility. Those accounts are anecdotes; Google's stated consumer rollout is to Ultra subscribers.

The specialized Carbon for cyber remains available through applications to the Fairwind Program. Scientists can apply for six months of free Antigravity access to Argon and Carbon, subject to eligibility, capacity, and selection.

I want to see how much of the document-to-model work a team can take on before a person has to intervene. Google's architectural examples make that a concrete experiment: give the agents plans, inspect what they build, and compare it with the originals. Broader access lets more people try it.
