TL;DR
Google has opened Gemini 4 Carbon to paying developers and is previewing /swarm, a system for coordinating coding agents. Headline token prices are temporary; the important question is whether parallel workers, verification and retries lower the cost of a result a human will actually accept.
Introduction
On November 10, Google expanded access to Gemini 4 Carbon after an initial release to cybersecurity defenders. In parallel, Antigravity introduced a preview of /swarm, which coordinates specialized agents working on a shared task. This is more than a model launch: it packages an organizational pattern for software development, with explicit roles for planning, implementation and checking. (Sources : Google — Gemini 4 Carbon is now available; Antigravity — Gemini 4 Carbon and /swarm)
The original Carbon release left questions about production pricing and the practical value of its scores. Our earlier analysis examined that initial announcement. The wider rollout gives buyers a more concrete basis for testing, not proof that their own workflows will improve.
What the swarm actually does
Antigravity says a coordinator launches parallel workers in isolated Git worktrees, while adversarial verifiers run tests and challenge changes. Agents communicate on a shared append-only message board. The interface exposes agent activity and a replayable graph, giving a reviewer a way to inspect how work progressed. (Source : Antigravity — Gemini 4 Carbon and /swarm)
Isolation limits workers editing the same working tree; it does not eliminate semantic conflicts when their changes are combined. An independently passing test suite can still miss a broken assumption between components. The meaningful human review point is therefore the integrated diff and its system-level tests, not simply a green check from each worker.
Google recommends Carbon for architecture, coordination and verification, with lighter models handling parallel implementation. That division resembles the planner-and-worker approach in our Muse Ember analysis, but it introduces coordination overhead: every handoff, repeated context and verifier call must earn its place. (Source : Google AI Studio — Gemini 4 Carbon developer guide)
Token pricing is not the price of a successful task
Google lists introductory Gemini API prices of $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, the listed rates become $4 and $20 respectively. The developer guide lists a 95% discount for cached input. (Source : Google AI Studio — Gemini 4 Carbon developer guide)
As an illustration, a direct API workflow consuming one million uncached input tokens and 100,000 output tokens would cost $3 at the introductory rates and $6 at the later rates. That is a token-only calculation, not an Antigravity subscription quote; actual usage, caching eligibility, tools, retries and other charges can change the total.
Google says idle agents park at zero token cost, but also warns that parallel multi-agent systems consume quota much faster than a standard chat. Zero idle-token use does not make active workers free. Compare a single-agent baseline with a swarm on the same tasks, then measure total token use, elapsed time, failed attempts and human review. Divide total cost by accepted results, not by attempted runs. (Source : Antigravity — Gemini 4 Carbon and /swarm)
Benchmarks and context: useful signals, not guarantees
Google reports Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than two times lower cost per task than the next best external model. Google also reports a second-place result on the Multimodal Knowledge Agent benchmark, developed with Artificial Analysis, at more than six times lower average cost per task than the leading model. These are vendor-reported comparisons: inspect benchmark settings, baselines and pricing before transferring the result to your workload. (Source : Google — Gemini 4 Carbon is now available)
The Gemini API provides a unified one-million-token context window shared by input and output. A large window helps fit project context, but capacity is not the same as retrieval accuracy, and repeated long prompts can become expensive. Set task-specific context and spending limits; test whether a smaller retrieved excerpt produces the same verified outcome. (Source : Google AI Studio — Gemini 4 Carbon developer guide)
Deployment boundaries
Google AI Ultra subscribers can use Carbon in Antigravity, while /swarm is rolling out in preview in the Antigravity app and CLI. Paid API availability and an Ultra subscription are different access and billing paths; do not infer a swarm run’s cost from API list prices alone. (Sources : Antigravity — Gemini 4 Carbon and /swarm; The Verge — Google releases Gemini 4 Carbon)
Start with reversible repository changes in an isolated environment. Restrict credentials and network access, require approval before deployment or external writes, and review tests for coverage rather than treating an agent verifier as an independent security guarantee. The opportunity is faster verified work; the risk is multiplying unreviewed actions just as quickly.
FAQ
Is /swarm generally available?
No. Antigravity describes it as a preview rolling out in its app and CLI.
Does Carbon’s introductory price last indefinitely?
No. The published Gemini API rates increase after November 30, 2026; check current terms before budgeting.
Does a million-token window guarantee accurate answers?
No. It is a capacity limit, not a measurement of retrieval quality or correctness.
Should every task use a multi-agent team?
No. Coordination and verification add overhead. Compare against a simpler baseline on the same acceptance criteria.