Gemini 4 Carbon

Learning Objectives

• Understand what Gemini 4 Carbon is and how Google positions it against Gemini 4 Argon
• Know who can use Carbon, what the API costs, and why Pro subscribers do not get it
• Read Google's benchmark and safety claims for what they are: mostly vendor-reported, with one outside benchmark so far

What Is Gemini 4 Carbon?

Gemini 4 Carbon is the second model in Google DeepMind's Gemini 4 generation, after Gemini 4 Argon. Google calls it its strongest model and built it for the work that tests a model's limits: long, multi-step coding and debugging, open-ended scientific research, and finding and patching software vulnerabilities.

Carbon does not replace Argon. Google says Argon remains the right choice for the majority of software engineering and knowledge work, while Carbon spends more tokens on a task to explore it thoroughly, reason through it and run longer checks on its own output. That makes Carbon slower and more expensive per task on easy work, and better on hard work. Choose by the difficulty of the job, not by which model is newer.

From Gated Preview to General Release

Carbon followed the same path as Argon. Google announced it on October 28, 2026 and gave it first to cyber defenders in its Fairwind Program, saying it would reach everyone else "as soon as possible" without naming a date. Thirteen days later, on November 10, it opened the model to paying developers, enterprises and Ultra subscribers.

The same day Google previewed a new swarm feature in Antigravity, its coding environment, which turns one prompt into a coordinated team of agents you can watch in real time. Google suggests giving Carbon the roles that need the deepest reasoning, such as architecture, coordination and auditing, and pairing it with faster, cheaper models for the rest of the work, with separate verifier agents testing each change.

Benchmarks and Early Results

Most figures below are reported by Google, including the ones run on other companies' benchmarks.

Google says Carbon took first place on CursorBench 4.0, Perplexity's WANDR test and Zapier's AutomationBench at less than half the cost per task of the next best model from another company. On the new Multimodal Knowledge Agent benchmark, which Google built with Artificial Analysis, it ranks second, at less than one-sixth of the leader's cost per task.

The first outside read came on November 2, 2026 from GDPx, a benchmark of hard agentic tasks across eleven fields published by the research firm SemiAnalysis. It called Carbon a real improvement and the only model in the same range as Claude Opus 5.5 and GPT-6 Astra. GDPx ranked Carbon first on biology, which SemiAnalysis attributes partly to lighter safety classifiers, and gave it a narrow win in software engineering, but Opus 5.5 still came out as the best all-around model.

Google's research examples are its own accounts. Its researchers used Carbon with an agent framework called AlphaProof Code to write millions of lines of formal proofs in Lean, a language for machine-checked mathematics, and prove key security properties of the actual code behind the Signal messaging protocol in weeks, work Google says normally takes expert teams years. Google also says that across seven protein families from a blind structure-prediction test, Carbon agents found forty times as many distant evolutionary relatives on average, which could give structure-prediction models more to work with. Google adds that broader benchmarking is still underway.

Safety and Security

Because Carbon is stronger at cyber work, Google says it tightened the safeguards it built for Argon. In simulated attack environments, it reports that Carbon blocked 99 percent of attack attempts from an adaptive red-teaming agent, cutting the breach rate per attempt to 0.8 percent from 3.1 percent on Argon. In outside testing by Gray Swan, automated cyberattack queries succeeded 2.17 percent of the time against 9.86 percent on Argon, and Trajectory Labs found no new universal jailbreaks in more than 13,000 requests.

Google also reports that Carbon passed 92 percent of a 375-task honesty test for agents, against 70 percent for Argon, mostly by declining to invent missing experimental data. At the October preview it said Carbon was the first model to let through no attacks at all on Gray Swan's indirect prompt injection benchmark. Google itself says prompt injection is not solved and that its own red-teaming still finds attacks that work in some cases.

Pricing & Access

Google launched Carbon at the same introductory API price as Argon, whose introductory rate ends on November 30, 2026 before doubling to 4 dollars and 20 dollars per million tokens. At the October preview Google said Carbon would follow the same path, so budget for the higher rate on anything you plan to run into December. Because Carbon uses more tokens per task than Argon, the same price per token still means a larger bill per job.

Strengths

• Hardest coding and research work: Google built Carbon for long, multi-step tasks and reports first place on CursorBench 4.0 and AutomationBench
• Security work: trained to find, exploit and patch vulnerabilities, with a separate version for vetted defenders
• Agent teams: the Antigravity swarm preview is designed around Carbon handling the hardest roles while lighter models do the rest
• Introductory price: the API rate matches Argon and sits well below other closed frontier models, at least through November 30
• Measured safety gains over Argon: lower breach and jailbreak rates in Google's and outside testers' cyber tests

Limitations & Considerations

• Not the best all-around model on outside tests: SemiAnalysis's GDPx put it in the same range as Claude Opus 5.5 and GPT-6 Astra, but rated Opus 5.5 best overall
• Narrow lead over Argon: Ars Technica found Carbon's lead over Argon slim on every test Google published at the preview, so for most work the cheaper-per-task Argon is the sensible default, as Google itself says
• Most numbers are vendor-reported: Google ran or reported nearly every figure on this page, including results on other companies' benchmarks
• Ultra only for consumers: Pro subscribers and free users cannot use it in the Gemini app or AI Mode
• More tokens per task: Carbon is designed to spend more tokens than Argon, so costs rise even at the same price per token
• Two versions: the vulnerability-hunting scores come from the Fairwind cyber version, which may behave differently from the public model you can call

Related Tools

• Gemini 4 Argon — the first Gemini 4 model, which Google still recommends for most work
• Gemini — Google's consumer app, where Carbon reaches Google AI Ultra subscribers
• Google Antigravity — Google's agent-first coding environment, where the swarm preview runs
• Claude Opus 5.5 — the model SemiAnalysis's GDPx rated best overall
• GPT-6 Astra — OpenAI's flagship, in the same range as Carbon on GDPx

Getting Started

1. Decide whether the task is hard enough to need Carbon; for routine coding and knowledge work, start with Gemini 4 Argon
2. Developers: select Gemini 4 Carbon on the paid API tier in Google AI Studio, or in Gemini for business
3. Consumers: subscribe to Google AI Ultra, then pick Carbon in the Gemini app or AI Mode in Search
4. In Antigravity, try the swarm preview on a large task and assign Carbon the coordinating and auditing roles
5. Check the price before December: the introductory API rate ends on November 30, 2026

Key Takeaways

• Gemini 4 Carbon is Google DeepMind's strongest model, previewed on October 28, 2026 for cyber defenders and opened on November 10 to paying developers, enterprises and Google AI Ultra subscribers
• Google says Gemini 4 Argon remains the better choice for most work, and Carbon is for the hardest coding, research and security tasks
• The API costs 2 dollars per million input tokens and 10 dollars per million output at the introductory rate, the same as Argon, and Carbon uses more tokens per task
• Google AI Pro subscribers do not get Carbon, and free users get neither Gemini 4 model
• Most benchmarks are Google's own; the outside GDPx test put Carbon level with Claude Opus 5.5 and GPT-6 Astra but rated Opus 5.5 best overall
• A new swarm preview in Antigravity turns one prompt into a team of agents, with Carbon intended for the hardest roles