Gemini 4 Carbon Now Generally Available: Paid Gemini API Developers and Ultra Subscribers Get Access Today at an Introductory $2 / $10 per Million Tokens, as Antigravity Previews Multi-Agent /swarm

Core Background and Industry Pain Points

When Google announced Gemini 4 Carbon on October 28, it went only to a set of cybersecurity defenders in the Fairwind Program, so developers and regular users could not use it. For teams that want a frontier model for long-running agent work, a model they can read about but not call is the main problem. On November 10, Google opened Carbon broadly, and Sundar Pichai, Logan Kilpatrick, and Josh Woodward confirmed the rollout on X.

Architecture Highlights and Underlying Mechanisms

Google positions Carbon one step above Argon. Argon remains the right choice for most coding and knowledge work, while Carbon spends more tokens to explore, reason deeply, and run longer self-verification loops, which suits multi-step coding and debugging and open-ended research. Alongside Carbon, Antigravity is previewing the /swarm slash command in the desktop app and CLI (official docs say it is available on Google AI Pro and Ultra). Your current conversation becomes the coordinator, which spins up named workers and, when several of them edit at once, can give each one an isolated git worktree; it usually adds an adversarial @verifier for every two to four workers. Everyone coordinates on a shared, append-only message board with @mentions, and idle members use no tokens. A side pane provides a Live Board, a Members roster, and a Graph timeline; in the CLI you can open a full-screen board with /swarm board or Alt+S, and you can pin a model to an agent. Google recommends Carbon for coordinator, architecture, cybersecurity auditing, and verification roles, and lighter, faster models for parallel file edits. It said Google engineers dogfooded /swarm internally for about five weeks on multi-package migrations, cross-subsystem investigations, and architectural redesigns, and warned that parallel multi-agent runs consume quota much faster than a standard chat.

Authoritative Benchmarks and Measured Comparisons

These figures are cited in Google's blog. Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, at more than 2x lower cost per task than the next best external model. On the Multimodal Knowledge Agent (MKA) benchmark launched with Artificial Analysis, it ranks #2 at more than 6x lower average cost per task than the leader. On safety, it blocked 99% of attack trajectories in simulated cyberattack environments, cutting the per-attempt breach rate from 3.1% on Argon to 0.8%. Gray Swan measured a 2.17% automated cyberattack query success rate (9.86% for Argon), and Carbon passed 92% of a 375-task agentic honesty benchmark (70% for Argon). Google researchers also used Carbon with a new framework, AlphaProof Code, to generate millions of lines of Lean proof code and prove key security properties of the actual code behind Signal's cryptographic protocol in weeks. Cursor then officially said Carbon is in the editor and updated its public CursorBench 4.0 leaderboard: Carbon High ranks #1 at 60.4%, with Medium at 60.1% and Low at 58.6%, just ahead of Fable 5.2 Max at 60.0%. At published pricing, Carbon High costs about $9.22 per task versus about $32.79 for Fable 5.2 Max. Perplexity also said on its official X account that Carbon scores 0.857 on its WANDR benchmark at $5.86 per task, matching Claude Fable 5.2 at about 40% of the cost (vendor figures).

Developer Adoption and Getting Started

Paying developers can call Carbon now through the Gemini API in Google AI Studio or Android Studio, and enterprises can pick it in Gemini for business through the API or the model drop-down. Ultra subscribers can use it directly in the Gemini app, AI Mode in Search, and Antigravity. Carbon is also live in Cursor. Cursor's docs say it has three effort levels (high, medium as the default, and low); as a third-party model it is not included in Cursor plan usage, and is billed from prepaid credit on The X Subscription or the Other Models pool on legacy plans; from December 1, cache reads are $0.20 per million tokens. It is generally available in GitHub Copilot for Pro+, Max, Business, and Enterprise at provider list pricing, and in Perplexity Computer for Pro and Max subscribers. The introductory price is $2 / $10 per million input / output tokens and becomes $4 / $20 after November 30, so teams planning heavy long-running workloads should finish their evaluation before the end of the month. Following Google's own guidance, keep Argon for everyday tasks and send the hardest debugging and architecture reviews to Carbon. In the Antigravity app or CLI, type /swarm, then describe the task, to try the preview (Pro and Ultra); in the CLI, open the board with /swarm board or Alt+S, and watch quota, because parallel runs consume it faster. Scientists and researchers can apply for a chance to get Antigravity's highest usage tier free for six months, including Argon and Carbon. The rankings are cited by Google, so retest cost and quality on your own tasks before switching.