--- title: "Google Unleashes Gemini 4 Carbon, Its Most Powerful AI Model" date: 2026-11-10 author: "Barry Elad" featured_image: "https://sqmagazine.co.uk/wp-content/uploads/2026/11/google-gemini-4-carbon-launched.jpg" categories: - name: "Artificial Intelligence" url: "/artificial-intelligence.md" tags: - name: "News" url: "/tag/news.md" --- # Google Unleashes Gemini 4 Carbon, Its Most Powerful AI Model Google released Gemini 4 Carbon on 10th November, 2026, which it calls its strongest model. Paying developers and enterprises get it through the API at $2 per million input tokens and $10 per million output tokens. ## The Brief - Gemini 4 Carbon keeps the same introductory price as Gemini 4 Argon, and it uses more tokens on hard problems. - 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. - Google AI Ultra subscribers are getting Carbon in the Gemini app, AI Mode in Search and Google Antigravity. - Argon stays Google’s pick for most software and knowledge work, while Carbon targets harder multi-step coding and scientific research. ## Carbon widens access after a two-week cyber preview Google first gave Carbon to [cybersecurity defenders](https://sqmagazine.co.uk/cybersecurity-job-statistics/) through its Fairwind Program two weeks ago. Today’s release follows the recent launch of Gemini 4 Argon, and Google still points most software engineering and knowledge work to that model. Carbon is built for “**tasks that test the limits of a model’s intelligence**,” Google said in its [announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/). Complex multi-step coding, debugging and open-ended scientific research fit that description. Carbon spends more tokens than Argon to reason deeply and run extended verification loops. A hard task can therefore cost more than the per-token price suggests. Paying developers can build with the **Gemini API in AI Studio and Android Studio**. Enterprises can use Gemini for business through the API or the model dropdown, or reach Carbon in Google Antigravity or CodeMender. > Gemini 4 Carbon is now available. It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning. If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems. > > — Google (@Google) [November 10, 2026](https://x.com/Google/status/2120186185423159331?ref_src=twsrc%5Etfw) ## Google pairs benchmark wins with cost claims Google reports that Carbon ranks **\#2** on the new Multimodal Knowledge Agent benchmark. Google built it with Artificial Analysis to test agents on dense files full of charts, images, blueprints, schematics and tables. Average cost per task runs more than **6X** lower than the top model’s. In pre-release tests, Carbon also ranked first among **23** frontier model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark. For engineering, [Google](https://sqmagazine.co.uk/google-gemini-ai-statistics/) says Carbon handles multi-package migrations and cross-subsystem root-cause investigations. A new /swarm preview in Google Antigravity turns one prompt into a team of agents that developers can track in real time. Developers can assign Carbon roles that demand the deepest reasoning, like architectural design and auditing. Faster, lighter models then work in parallel isolated workspaces, with adversarial verifiers testing every change. Anyone comparing [AI coding tools](https://sqmagazine.co.uk/best-ai-coding-tools/) now has another multi-agent option to test. **Google’s researchers also put Carbon to work on science**. Alongside other models, Carbon helped AlphaProof Code agents generate millions of lines of Lean. That let them prove key security properties of the code behind Signal’s cryptographic protocol in weeks. Google says that task historically takes expert teams years. For protein structure prediction, Carbon increased the homologs retrieved by **40-fold** on average across 7 low-homolog CASP17 families. Google says broader benchmarking is still underway. ## Cyber results come from testers inside and outside Google Through the Fairwind Program, Carbon leads on **SEC-bench Pro** for finding vulnerabilities, **ExploitGym** for exploiting them and **CWE-bench v1** for patching them. In Project Fortify, Google’s autonomous system for scanning, triage and remediation, Carbon hit **100%** recall on scanning and **95%** precision on triage. That came on an internal benchmark. In open-source repository testing, recall rose more than **25%** relative to Argon. Google’s Cloud Vulnerability Research team used Carbon to build a proof of concept for two bugs in popular open-source software. It showed both were exploitable, comparable in severity to the recent [PHP2Shell exploit](https://sqmagazine.co.uk/php-critical-security-update-two-9-2-flaws/). That helped maintainers reclassify the bug as critical and issue a patch. In Wiz’s Scan for Good program, the team used Carbon on two major SaaS providers. It found 0-day vulnerabilities in under two hours, compressing weeks of research into one afternoon. Gaps remain. The announcement names no CVE IDs, open-source project or SaaS providers. Google said the research team’s details will follow once responsible disclosure ends. OpenAI has also expanded its [Daybreak cybersecurity AI](https://sqmagazine.co.uk/openai-expands-daybreak-powerful-cybersecurity-ai/) offering. In Google’s simulated cyberattack tests, Carbon blocked **99%** of attack trajectories from an adaptive red-teaming agent. The per-attempt breach rate fell to **0.8%** from **3.1%** on Argon. On Gray Swan’s indirect prompt injection benchmark, Carbon reached a 0% attack success rate, a first. Google says its own red teaming shows models can still improve against sophisticated attacks in some distributions. ## What’s Next? Teams already running Argon can start with Google’s updated **Gemini 4 developer guide**, which covers migrating workflows. Scientists can still apply for free access to Argon and Carbon in Antigravity’s highest usage tier for six months. Onboarding runs on a weekly rolling basis, subject to eligibility, capacity and selection. Cybersecurity defenders can apply to the Fairwind Program for Carbon for cyber, also on a rolling basis. Definition of Prompt Injection. Link to full glossary entry follows the description.**Prompt Injection**Prompt injection is an attack in which crafted input makes a large language model follow the attacker's instructions instead of its operator's. [Read more about Prompt Injection](https://sqmagazine.co.uk/glossary/prompt-injection/) Definition of Frontier Model. Link to full glossary entry follows the description.**Frontier Model**A frontier model is a highly capable [general-purpose AI](https://sqmagazine.co.uk/glossary/general-purpose-ai/) model that matches or exceeds today's most advanced systems, and triggers safety obligations. [Read more about Frontier Model](https://sqmagazine.co.uk/glossary/frontier-model/)