---
type: "post"
id: "gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54"
slug: "gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54"
title: "Gemini 4 Carbon ranks #1 on CursorBench 4.0 and Perplexity's WANDR benchmarks for deep agentic work."
description: "Google says Gemini 4 Carbon is rolling out in Antigravity alongside a preview of /swarm, its native multi-agent orchestration mode."
retrieval_nugget: "Google says Gemini 4 Carbon is rolling out in Antigravity alongside a preview of /swarm, its native multi-agent orchestration mode. The combination is aimed at long, complex work: parallel agents can build in isolated worktrees while dedicated verifiers test and challenge their changes. Access is limited, and Google warns that swarms consume quota faster."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: ["google","gemini"]
entities: ["Google","Google Antigravity","Gemini 4 Carbon","/swarm","CursorBench 4.0","WANDR"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Gemini 4 Carbon now ranks first on CursorBench 4.0 and Perplexity’s WANDR benchmarks for deep agentic work."
prior_refs: ["https://newruntime.com/posts/gemini-4-carbon-becomes-1-on-artificial-analysis-8055b2/"]
source_url: "https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity"
source_urls: ["https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54/","markdown":"https://newruntime.com/posts/gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54.md","json":"https://newruntime.com/posts/gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54.json"}
markdown_url: "https://newruntime.com/posts/gemini-4-carbon-ranks-1-on-cursorbench-4-b4ad54.md"
---

# Gemini 4 Carbon ranks #1 on CursorBench 4.0 and Perplexity's WANDR benchmarks for deep agentic work.

## Retrieval answer

Google says Gemini 4 Carbon is rolling out in Antigravity alongside a preview of /swarm, its native multi-agent orchestration mode. The combination is aimed at long, complex work: parallel agents can build in isolated worktrees while dedicated verifiers test and challenge their changes. Access is limited, and Google warns that swarms consume quota faster.

## Summary
Google says Gemini 4 Carbon is rolling out in Antigravity alongside a preview of /swarm, its native multi-agent orchestration mode. The combination is aimed at long, complex work: parallel agents can build in isolated worktrees while dedicated verifiers test and challenge their changes. Access is limited, and Google warns that swarms consume quota faster.

## What happened
The Antigravity Team announced the rollout on November 10, 2026. Gemini 4 Carbon is being added to Google Antigravity, paired with /swarm in preview. Google describes Carbon as its strongest model for long-horizon software engineering, knowledge work, scientific discovery and cybersecurity. It says Carbon ranks first on CursorBench 4.0 and Perplexity’s WANDR, without publishing scores in this announcement.

Google presents /swarm as an alternative to using one agent for a scoped task or a primary agent that delegates isolated subtasks to helpers. In its description, /swarm passes a prompt to a group coordinated by a lead agent. That coordinator starts specialized workers in isolated git worktrees, so their edits do not collide, and includes adversarial verifiers tasked with running tests and challenging diffs. Agents communicate through a shared, append-only message board. Google says agents waiting for input incur zero token cost and activate when another agent needs them.

Antigravity includes a Swarm UI in its side pane, with a live Board feed, a Members roster and an interactive Graph timeline. Users can inspect tool activity, open an agent’s trajectory and steer it, or replay how the team moved from brief to verified code. In the CLI, users can open a live terminal board with `/swarm board` or Alt+S.

Google says teams can pin different models to different agents. Its suggested division of labor is to use Carbon for coordinator, architecture, cybersecurity audit and adversarial verification roles, while faster, lighter models handle parallel file edits and routine implementation. Google reports five weeks of internal use by its engineers on multi-package migrations, cross-subsystem root-cause investigations and architectural redesigns. The company cautions that parallel multi-agent systems consume quota much faster than standard chat, even with zero-token idle waits and dynamic model switching.

Google AI Ultra subscribers can use Gemini 4 Carbon in the Antigravity App, CLI and IDE extensions. /swarm is rolling out in preview across the Antigravity App and CLI. Enterprises can access Carbon in Gemini for business. Scientists and researchers can apply for a chance to access Argon and Carbon through Antigravity’s highest usage tier free for six months.

## What the archive adds
New Runtime’s earlier coverage says Gemini 4 Carbon became generally available in GitHub Copilot on November 10, 2026 ([earlier note](https://newruntime.com/posts/gemini-4-carbon-in-github-copilot-44e8c2/)), and became available in Devin Desktop and CLI on November 11 ([earlier note](https://newruntime.com/posts/gemini-4-carbon-is-live-in-devin-desktop-f5900a/)). Those were access changes; this is a new product event, adding Carbon to Antigravity and introducing /swarm.

## Trend signal
This supports **Harness architecture outlives model choice**: [the pattern](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/) argues that the production harness is a more durable boundary than the model. Antigravity’s role-specific model pinning and orchestration make that idea concrete, though the announcement does not establish how well the setup works outside Google’s internal use.

## Why now
Google says its engineers spent five weeks dogfooding /swarm before this preview. The announcement frames the feature as an answer to tasks spanning dozens of files and test suites, where a single agent or coordinator can become a bottleneck.

## Why it matters
For builders, the practical proposition is not simply “more agents,” but a workflow that separates parallel implementation from verification and makes coordination visible. Isolated worktrees can reduce edit collisions; explicit verifiers and a replayable graph can make it easier to inspect how changes were produced. Those are useful properties for migrations or investigations that span multiple packages, but they do not by themselves establish correctness.

The trade-off is operational. More agents can create more work to review and consume quota substantially faster, according to Google. A team evaluating /swarm should therefore compare completed, verified work—not the number of parallel tasks—with the cost and review time of its current process. Carbon’s benchmark position may help motivate a trial, but Google’s announcement gives no scores or independent workflow evaluation. Access also depends on Google AI Ultra or business availability, limiting who can test the setup now.

## Limits and uncertainty
Google’s announcement does not provide scores for CursorBench 4.0 or WANDR, or an independent comparison of Carbon against other models. The internal dogfooding account gives no task counts, completion rates, measured time savings, defect rates or quota totals. It remains unclear how /swarm behaves on less controlled repositories, how much human review the verifier agents require, and when the preview will become generally available. The feature’s stated zero-token idle cost does not remove the broader quota cost of parallel execution.

## Skill to build now
For a task that already spans multiple packages, define separate roles before trying a swarm: one coordinator, isolated implementation workers and a verifier with explicit tests and acceptance criteria. Keep the task small enough to compare against a single-agent run. Record quota consumed, human review time, test results and collisions; these are the measures needed to decide whether parallel orchestration helped.

## Reader takeaway
Antigravity’s /swarm preview combines parallel workers, isolated worktrees, dedicated verifiers and a visible coordination surface. It is worth evaluating for broad engineering tasks if you can access it, but Google itself warns that parallel work burns quota faster. Treat benchmark claims and internal examples as reasons to test—not proof that swarming improves your team’s results.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)