What Is Gemini 4 Carbon? Features, Uses, and Capabilities

Wiz’s security team used Gemini 4 Carbon to scan two major SaaS providers and found zero-day vulnerabilities in under two hours. That kind of research usually takes weeks.

Google built Carbon for long, multi-step work like this. It is Google’s AI model for tasks such as finding security issues in code. Google DeepMind announced it on October 28, 2026, and made it available to all paying developers on November 10. Using Artificial Analysis’ Intelligence Index v5.0.1, this guide covers Gemini 4 Carbon’s benchmarks, pricing, coding, cybersecurity, and availability.

Data checked on November 11, 2026. Rankings, prices, and access can change, so confirm them on the linked sources before you buy or build.

What Is Gemini 4 Carbon?

Gemini 4 Carbon is a proprietary reasoning model from Google DeepMind. It accepts text, image, audio, and video input and returns text output. Google’s Cloud documentation describes it as a frontier model for deep work in coding, finance, legal, cybersecurity, and science.

• Model type: A reasoning model in the Gemini 4 family, released after Gemini 4 Argon.
• Main purpose: Hard, multi-step work that tests the limits of a model’s intelligence.
• Target workloads: Complex coding and debugging, security research, scientific discovery, and analysis of dense documents.
• Reasoning style: It uses extra tokens to explore a problem, reason deeply, and run extended checks before answering.

Key Features of Gemini 4 Carbon

Carbon’s features group into five areas: coding, vulnerability detection, code analysis, long-running agent work, and reasoning.

Advanced Coding and Software Engineering

Carbon builds full-stack apps and checks its work as it goes. During web development, it builds and tests the app, fixes errors, and checks both the code and how the app looks on screen.

Output quality shows up in public rankings. At launch, Carbon gained 89 Elo points on WebDev Arena and landed at #4 with a score of 1768.

Google also reports a #1 rank on CursorBench 4.0. Its cost per task is more than 2X lower than the next best external model.

Design skills improved too. In Google Antigravity, Google’s agentic development platform, Carbon redesigned the Google Merchandise Store in two prompts. It proposed five art direction concepts and generated new product images with Nano Banana 2.1.

Carbon also builds games. In Antigravity, it built an open-world 3D game of Google’s Mountain View campus. Subagents modeled assets in Blender, and the browser game runs on Three.js with WebGL shaders.

Also Read: Best AI models for coding for how Carbon ranks against the rest.

Cybersecurity and Vulnerability Detection

Carbon handles every stage of security testing. It finds security flaws, checks whether attackers can use them, and fixes the root cause. Google tested it without guardrails on three external benchmarks and says it performed better than other models on all three.

Source: Google, “Gemini 4 Carbon: our next frontier model” (October 28, 2026)

Data checked on November 11, 2026.

The exploit score helps teams decide which security issues to fix first. If attackers cannot use a flaw, it may not need immediate attention. But if an exploit works, it becomes a priority.

Code Analysis and Security Testing

Carbon writes safer code and finds more real bugs than Argon. On Google’s internal benchmarks, its success rate at generating secure code is more than double Argon’s.

Google’s Chrome team reports that Carbon produced high-quality patches more than 10X better than another lab’s large frontier model. Wiz, a cloud security platform, also saw Carbon lead its web penetration testing benchmarks.

Project Fortify, Google’s system for scanning, triage, and remediation, gives the deepest numbers. Carbon hit 100% recall on scanning and 95% precision on triage on Google’s internal benchmark. On open-source repositories, recall rose more than 25% over Argon, and Carbon surfaced real logic flaws that nobody had detected.

Long-Running Agentic Tasks

Google built Carbon to handle long-running tasks. Its teams tested it for several days and saw it work through difficult problems. It also found the root causes of hidden bugs and fixed them across large codebases and low-level systems.

Staying on task matters as much as endurance. Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts at 0.4%. Google also reports 26% less overreach and 29% less under-service, so Carbon stays inside project limits.

In live coding environments, serious misalignment alerts dropped by almost 80%.

In a benchmark with 375 agentic honesty tasks, Carbon completed 92% of them, compared to 70% for Argon. When a request is unclear, it checks available sources and points out missing information before taking action.

Advanced Reasoning Capabilities

Carbon can also solve problems in science, math, and finance, not just coding. Google reports improvements in physics, math, and life sciences, including testing ideas and solving multi-step problems.

• Drug research: Google reports that Carbon is now the leading model for Axiom Bio, a drug toxicity research company. It found toxicity mechanisms the company had missed.
• Protein science: On average, Carbon retrieved 40 times more homologs across seven low-homolog CASP17 protein families.
• Lab data: Given only notebooks and raw data, Carbon diagnosed problems like incomplete automated mixing in deep wells.

Across legal, finance, and enterprise work, Carbon follows strict schemas and formatting to produce polished spreadsheets, presentations, and reports. In finance, it ranked first among 23 frontier model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark.

Google built the Multimodal Knowledge Agent benchmark with Artificial Analysis. Carbon ranks #2 there, at more than 6X the top model's average cost per task.

What Makes Gemini 4 Carbon Different?

Carbon trades speed for depth. Google calls Argon the better choice for most software engineering and knowledge work. Carbon targets tasks that test the limits of a model’s intelligence.

That deeper approach can be seen in three ways:

• More thinking for each task: Carbon uses more tokens than Argon to explore ideas, think through problems, and check its own work.
• One shared token limit: Carbon has a single limit of 1,048,576 tokens for both input and output. Earlier Gemini models used a different approach.
• Different controls: Carbon lacks the MINIMAL thinking level, so developers moving from Gemini 3.1 Pro must change that setting.

What Can You Use Gemini 4 Carbon For?

Carbon fits work that is long, risky, or spread across many files. Four uses stand out.

Finding and Fixing Software Vulnerabilities

Start with a repository and let Carbon work through four steps:

• Scan the code for flaws.
• Build a proof of concept that shows whether each flaw is exploitable.
• Draft a patch that fixes the root cause.
• Review and test the patch before you merge it.

Carbon supports structured output, so you can request findings in a fixed format for your issue tracker.

Google’s Cloud Vulnerability Research team used Carbon to prove that two bugs in popular open-source software were exploitable. The severity matched the recent PHP2Shell exploit. Maintainers reclassified the bug as critical and issued a patch.

Analyzing Large Codebases

Carbon can process up to 1,048,576 tokens in one request, allowing it to handle large parts of a codebase at once. Google says it works well for tasks such as updating multiple packages and making complex changes to software architecture.

You can use Carbon to check how different parts of the code depend on each other, review design choices, and plan changes before editing the code. It also remembers the context automatically, so asking repeated questions about the same code costs less.

Testing and Securing Applications

Security teams use Carbon to test live software the way an attacker would. In Wiz’s Scan for Good program, the team used Carbon to assess public digital infrastructure. Scanning two major SaaS providers, it uncovered zero-day vulnerabilities in under two hours.

That compressed weeks of research into one afternoon. Wiz worked with the affected organizations to fix the flaws in record time. That cut the potential impact on tens of thousands of companies.

Securing code can also mean proving it correct. Google researchers built AlphaProof Code, an agentic framework for formal verification. Agents used Carbon, alongside other models, to generate millions of lines of Lean, a language for writing mathematical proofs.

The agents proved key security properties of Signal’s protocol code in weeks. That task historically takes expert teams years.

Automating Software Engineering Tasks

Google Antigravity is previewing /swarm, which turns one prompt into a coordinated team of agents you can track live. Developers give Carbon the roles that need the deepest reasoning: architectural design, coordination, and auditing.To build a lead-and-helpers setup of your own, a coding agent builder lets you set the roles, the review step, and where a person signs off. Faster, lighter models work in parallel, isolated workspaces, while adversarial verifiers test every change.

In one example, the Carbon team used seven publicly available City Council planning PDFs and one prompt to build an interactive 3D model of Google's Platform 37 office using Three.js. The model had 906K triangles and was built over 10 overnight cycles.

Starting a build from one prompt is the idea behind vibe coding. Here, the prompt starts a whole agent team.

Gemini 4 Carbon Specifications

Carbon takes up to 1,048,576 tokens per request and supports text, image, audio, and video input.

Model Details

Source: Google Cloud documentation, Gemini 4 Carbon

Data checked on November 11, 2026.

If your project needs model tuning or real-time Live API sessions, consider other Gemini models first.

Getting Started for Developers

Set the model ID to gemini-4-carbon in Google AI Studio or the Gemini API. Google directs developers to its updated Gemini 4 developer guide for best practices on migrating workflows to the model. Teams moving from Gemini 3.1 Pro can follow Google’s side-by-side migration guide.

Benchmark and Speed Data

Independent data shows Carbon ranks high on intelligence and low on speed and cost. Intelligence Index v5.0.1 combines nine evaluations, among them Terminal-Bench 4.0 and FrontierCode 2.0.

Source: Artificial Analysis, Gemini 4 Carbon release page

Data checked on November 11, 2026.

Artificial Analysis tracks three versions of Carbon, and each one leads a different metric.

Source: Artificial Analysis, Gemini 4 Carbon release page

Data checked on November 11, 2026.

Prices differ by up to 1.2x across the three versions. Artificial Analysis shows the standard rates of $4.00 and $20.00, not Google's introductory price.

Carbon ranks #4 in intelligence but #62 out of 72 in performance and #70 out of 73 in verbosity. It takes 66.01 seconds to produce its first answer token, making it unsuitable for live chat. Use Carbon for background tasks and review work instead.

The cost rank follows from verbosity. Carbon wrote 205M output tokens on the Intelligence Index.

Pricing

Carbon costs $2 per million input tokens and $10 per million output tokens until November 30, 2026. Cached input tokens cost 95% less than the input price.

Source: Google AI Studio

Data checked on November 11, 2026.

For example, processing 1M input tokens and 200K output tokens costs $4.00 at the introductory rates. After November 30, the same usage costs $8.00, before any caching discounts.

Who Can Access Gemini 4 Carbon?

Any paying Gemini API developer can use Carbon today. The cyber version stays limited to vetted defenders.

Google began with a small group of cyber defenders and widened access two weeks later. Carbon joins Argon across these products, so teams can match the model to the task.

Source: Google, “Gemini 4 Carbon is now available” (November 10, 2026)

Data checked on November 11, 2026.

The Fairwind Program gives approved security teams early access to Google's models designed for security work. Google works with more than 650 partners worldwide and checks applicant organizations before accepting them.

To apply, contact your Google Cloud account team. Approved projects use a separate model ID, gemini-4-carbon-cyber-permissive.

How Does Gemini 4 Carbon Help With Cybersecurity?

Carbon gives security teams a faster path from finding a flaw to fixing it, with controls that limit misuse. Google reports these safeguard results:

• Blocks attack attempts: In simulated cyberattack environments, Carbon blocked 99% of attack paths from an adaptive red-teaming agent. The per-attempt breach rate fell to 0.8%, from 3.1% on Argon.
• Resists automated attacks: Automated cyberattack queries succeeded 2.17% of the time in Gray Swan’s testing, against 9.86% for Argon.
• Resists external attacks: Trajectory Labs found no new universal jailbreaks after 132 hours of testing and more than 13,000 requests.
• Blocks hidden instructions: Prompt injection is when attackers hide commands in content that an AI reads. Carbon had a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection benchmark. Google says no other model had achieved this result before.

Google says stronger defensive reasoning must not become an offensive risk. That is why the cyber model sits behind a vetting program.

Fairwind partners are restricted to permitted dual-use tasks, including authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes. Malicious tasks, such as creating malware, are prohibited.

Google admits that prompt injection resistance still needs work. Its automated red teaming shows models can improve against sophisticated attacks in some cases.

Conclusion

Gemini 4 Carbon is the right pick for tasks that are long, risky, and spread across many files. Google made it for coding, cybersecurity, and scientific work that tests a model's limits.

It ranks #4 of 74 models on the Artificial Analysis Intelligence Index with a score of 60. Google also reports leading results in finding, exploiting, and patching software flaws.

It is slow and token-heavy: 78 tokens per second and $7.76 per Intelligence Index task. Keep faster models for everyday work.

You can use Carbon through Google AI Studio if you have paid access. The cybersecurity version is available only to those approved through the Fairwind Program. Google’s introductory rates are $2 and $10 per million tokens until November 30, 2026.

If you want to build an AI-powered app instead of testing a model, try Emergent’s AI app builder. Describe your idea in plain language, and Emergent builds the full-stack app: frontend, backend, database, auth, payments, and deployment. You own the code, the data, and the deployment.