Gemini 4 Carbon: Google's New AI Model and /swarm Explained

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google's newest frontier AI model, which Google calls its "strongest model", built for long-running agent work in software engineering, knowledge work, scientific reasoning and cybersecurity ^[1]. It was announced in a Google blog post by Koray Kavukcuoglu of Google DeepMind on 10 November 2026 ^[1].

Carbon sits alongside Gemini 4 Argon, which launched in October ^[2]. Google says Argon "remains the ideal choice for the majority of software engineering and knowledge work", while Carbon is meant for tasks that "test the limits of a model's intelligence", such as complex multi-step coding and debugging and open-ended scientific research ^[1]. Google adds that Carbon does this by "using more tokens than Argon" to explore, reason and run verification loops ^[1].

The Verge reports that Carbon was first released to a select group of "cybersecurity defenders" two weeks before the public launch ^[2]. Google's safety report describes Carbon as "the next model in our Gemini 4 series" ^[5].

How can you use Gemini 4 Carbon?

Developers and businesses can use Carbon through the Gemini API in Google AI Studio and Gemini for business; consumer access is through a Google AI Ultra subscription ^[1].

• Gemini app and AI Mode in Search: rolling out to all Google AI Ultra subscribers ^[1][4].
• Google Antigravity: Ultra subscribers can use Carbon in the Antigravity app, command-line tool and IDE extensions ^[3].
• Enterprises: access through Gemini for business ^[1][3].
• Scientists: Google says researchers can apply for free access to Argon and Carbon in Antigravity's highest usage tier for six months, subject to eligibility, capacity and selection ^[1].
• Security teams: defenders can apply to Google's Fairwind Program for access to a "Carbon for cyber" version on a rolling basis ^[1].

9to5Google notes that Argon stays available for AI Pro subscribers and above ^[4]. The Verge says it is not clear when, or whether, Carbon will reach lower tiers or free users ^[2].

How much does Gemini 4 Carbon cost?

Through 30 November 2026, the API price is $2 per million input tokens and $10 per million output tokens, the same introductory price as Argon; after that it rises to $4 and $20 ^[1]. A token is a small chunk of text, roughly part of a word, that AI models read and write.

The Verge calculated the introductory price at around a fifth of the cost of OpenAI's GPT-6 Astra and a quarter of Anthropic's Fable 5.2 per token ^[2]. It added an important caveat: models use different numbers of tokens for the same task, so a cheaper per-token price does not necessarily mean a cheaper job ^[2].

That caveat matters for Carbon in particular, because Google itself says it uses more tokens than Argon ^[1]. In one developer's published test of 20 coding runs, Carbon produced 1.82 million output tokens against 1.37 million for Argon, and its total bill at list price was similar to that of a model with four times the per-token price ^[6]. That is a single small test, not an independent benchmark, but it shows why per-token prices alone can mislead.

What is /swarm in Google Antigravity?

/swarm is a preview feature in Google Antigravity, Google's agent-based coding tool, that turns a single prompt "into a coordinated team of agents you can track in real time" ^[1].

The Antigravity team describes the setup: a coordinator agent starts specialised parallel workers in isolated git worktrees so their edits do not collide, while "adversarial verifiers" run tests and challenge changes ^[3]. Agents share an append-only message board and, Google says, sit idle at zero token cost until needed ^[3]. Each agent can be assigned a different model, so Carbon can take coordinator, architecture or auditing roles while lighter models handle routine edits ^[1][3].

The trade-off is usage. A developer testing Carbon reported that the /swarm option is labelled "Preview & Quota Heavy" in Antigravity's own menu, and that four parallel sessions hit the Ultra plan's quota after about two hours ^[6]. Running many agents at once multiplies token use. For background on how AI agents work and where they fail, see What Are AI Agents? How They Work, What They Can Do and Where They Fail.

How good is Gemini 4 Carbon on benchmarks?

Google says Carbon ranks first 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, but these rankings are reported by Google ^[1]. Google also says Carbon ranks second on a new Multimodal Knowledge Agent (MKA) benchmark that Google launched with Artificial Analysis ^[1].

For cybersecurity, Google says the Fairwind version of Carbon leads on SEC-bench Pro, ExploitGym and CWE-bench v1, which test finding, exploiting and patching software flaws ^[1]. Some figures, such as 100% recall on scanning, come from what Google calls "our internal benchmark" ^[1].

Independent context is limited so far. The Verge said benchmarks Google cited when it first announced Carbon suggest it is one of the most powerful models across most metrics, but behind Anthropic's Fable 5.2, and warned that companies "can and do cherry-pick benchmark results" ^[2]. A single developer's two-hour test found Carbon slightly ahead of Argon on code changes taken from a production monorepo (17 runs that passed every test out of 20, versus 16) but slow, averaging about 30 minutes per fix ^[6]. The Pouyam guide ChatGPT vs Gemini vs Claude in 2026: Models, Prices and Key Differences covers the wider field.

Is Gemini 4 Carbon safe?

Google's safety and security report says it cannot rule out that Carbon, without its extra safeguards, reaches the "Cyber Uplift Level 1" and biological "CBRN Uplift Level 1" Critical Capability Levels in its Frontier Safety Framework, and that it has deployed mitigations that lower the residual risk "to an acceptable level" ^[5]. These determinations were carried forward from Argon, because Google judged that Carbon did not show a material capability change ^[5].

Google reports, from its own tests, that Carbon blocked 99% of attack trajectories from its adaptive red-teaming agent in simulated environments (the safeguards lowered the per-attempt breach rate to 0.8%), and that Carbon passed 92% of a 375-task agentic honesty benchmark versus 70% for Argon ^[1]. It also cites external testing: Gray Swan's testing put the automated cyberattack query success rate at 2.17% (9.86% for Argon), and Trajectory Labs found no new universal jailbreaks across 132 hours ^[1].

Google says prompt injection, where hidden instructions in web pages or files hijack an AI agent, "remains an area for continued improvement" ^[1]. If you give an agent access to files or accounts, keep sensitive data out where possible; Is ChatGPT Safe for Your Privacy? What AI Chatbots Do With the Data You Share explains what chatbots do with what you share.

Should you switch to Gemini 4 Carbon?

Google itself recommends Argon for the majority of software engineering and knowledge work and Carbon for the hardest tasks, so most people do not need to switch ^[1].

• Consider Carbon for complex multi-step coding and debugging, open-ended scientific research or security auditing, where extra thinking time is worth it ^[1][3].
• Stay with Argon or a lighter model for most software engineering and knowledge work, as Google advises, especially if you watch costs or quotas ^[1][6].
• Plan for the price change: budgets set on the $2/$10 price will double from 1 December 2026 ^[1].
• Check results: strong benchmark scores do not prevent mistakes. AI Hallucinations Explained: Why Chatbots Make Things Up and How to Get Accurate Answers explains why AI models still make things up.

Until independent results are available, treat Google's rankings as claims and test Carbon on your own tasks before relying on it.