Gemini 4 Carbon is Google's newest artificial intelligence model, announced by Google DeepMind on 28 October 2026, just two weeks after the launch of Gemini 4 Argon. Google describes Gemini 4 Carbon as its strongest model to date, built especially for cybersecurity defence and for long-horizon agentic coding and engineering work, where an AI system carries out tasks over hours or days, such as finding and patching software bugs across large codebases. The announcement attracted attention because it positions Google directly in the race to use AI for defending computer systems, not only for writing text, answering questions or generating images.

This article explains Gemini 4 Carbon in plain terms. It sets out what Google claims the model can do, how it compares with its predecessor and with rivals on the benchmarks Google chose to publish, how access and pricing will work, what safeguards are in place, and why the claims should be read with caution. The facts come from Google DeepMind and Reuters reporting in October 2026. Benchmark figures are Google's own results, and independent verification is still pending. Gemini 4 Carbon is not publicly available to developers, businesses or consumers at the time of writing, and no date for wider release has been given.

Gemini 4 Carbon: What Google Announced

Gemini 4 Carbon was announced on 28 October 2026, two weeks after Gemini 4 Argon, which was presented as a general-purpose model. Carbon is described as a more specialised model, focused on two demanding areas. The first is cybersecurity defence: finding vulnerabilities in software, validating whether a weakness can actually be exploited, and fixing the root cause with a patch. The second is long-horizon agentic coding and engineering: tasks that run for a long time, involve many steps, and require the model to keep track of a large codebase, test its changes and correct mistakes. Google says Carbon is built to work on those tasks for days, rather than only answering a single question in one step.

The model has a context window of 1 million tokens. A token is a small unit of text or code, and the context window is how much material the model can consider at once. A 1-million-token window allows the model to work with very large codebases or long documents in a single task. For comparison, Gemini 4 Argon has a context window of 1,048,576 tokens. The difference in the headline numbers is small, but the two models are aimed at different work: Argon as a broad model, Carbon as a specialist for security and long engineering tasks. The distinction matters because a model that is strong at one kind of task is not automatically strong at every kind of task.

Background: Argon, Carbon and the Race for Stronger Models

Gemini 4 Carbon arrives in a fast-moving competition between the leading AI labs. Each new model is presented with benchmark scores that are meant to show progress over the previous generation and over rivals. Google, Anthropic, OpenAI and others release models in quick succession, and a lead on one benchmark can be overtaken within weeks. Argon was announced two weeks before Carbon, which shows how quickly the cycle is moving. Carbon's focus on cybersecurity is also a sign that AI companies are looking for practical, high-value uses beyond general chat, including work that security teams and software engineers do every day.

The wider context includes heavy investment in AI infrastructure, from chips and data centres to the models themselves. Readers can see that demand in the Microsoft and Alphabet earnings reported in October 2026, in the Anthropic IPO story, and in the launch of Gemini 4 Argon. Those stories show why companies are competing so hard: AI models are now central to search, cloud services, coding tools and security products. A model that can find and fix bugs reliably could save engineering time and reduce the risk of cyberattacks, but it could also be misused if it fell into the wrong hands, which is why Google is limiting initial access and testing safeguards.

What Happened: Benchmark Scores and Google's Claims

Google reported that Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index version 5, ranking first, one point ahead of Claude Opus 5.5. That index is a composite measure that brings together results from several tests, and a one-point lead is narrow. It suggests Carbon is competitive at the top of the field on the tests included, rather than far ahead of every rival. Readers should note that different benchmark suites rank models differently, and that a single index does not capture every capability, cost or reliability question that matters in real use.

On cybersecurity benchmarks, Google reported three headline figures, all run without guardrails. SEC-bench Pro, which tests finding vulnerabilities, was 88.2%. ExploitGym, which tests exploitation validation, was 54.8%. CWE-bench, which tests root-cause patching, was 71.0%. Google also reported internal test results: secure code generation at more than twice the success rate of Argon, and the Chrome team reporting more than 10 times as many high-quality patches as another laboratory's large model. Those are Google claims from its own testing, attributed as such, and they have not yet been independently verified. The fact that the cyber benchmarks were run without guardrails is important: guardrails are safety limits that can change how a model behaves, so scores without them may not match performance in a deployed product with safeguards in place. On a separate coding-related test, WebDev Arena, Carbon scored 1769 Elo, placing fourth, up 89 points.

Why Gemini 4 Carbon Matters for Cyber Defence

Gemini 4 Carbon matters because cybersecurity is a race between defenders and attackers. Software vulnerabilities are found every day, and the time between discovery and patching can decide whether systems are breached. A model that can find vulnerabilities, check whether they are real, and produce a correct patch could help defenders move faster and reduce the window of risk. For large organisations with huge codebases, that kind of help could be valuable, especially when security teams are short-staffed and the volume of code keeps growing. Long-horizon work matters here because patching a real system is rarely a one-step task; it involves understanding the code, making a change, testing it and checking for side effects.

The same capability also raises safety questions. A model that is good at finding and exploiting vulnerabilities could be used offensively if it were widely available without controls. That is why Google is starting with a limited group of cyber defenders, taking part in voluntary pre-release testing with the United States government, and continuing to test safeguards. The balance between helping defenders and limiting misuse is central to how Carbon will be released. Readers should therefore see the cyber scores not only as a technical achievement but as a reason for caution about who gets access first and under what conditions.

Key Facts: Access, Pricing and Limits

The following summary brings together the verified facts about access, pricing and the limits of the current claims. Gemini 4 Carbon is a new model, and several details remain to be confirmed as wider testing proceeds.

  • Announced by Google DeepMind on 28 October 2026, two weeks after Gemini 4 Argon; described by Google as its strongest model.
  • Built for cybersecurity defence and long-horizon agentic coding and engineering, including tasks running for days across large codebases.
  • Scores 59 on the Artificial Analysis Intelligence Index version 5, ranked number one, one point ahead of Claude Opus 5.5.
  • Cyber benchmarks reported by Google, run without guardrails: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%; secure code generation more than twice Argon, Chrome team reporting over 10 times more high-quality patches than another lab's large model.
  • WebDev Arena score 1769 Elo, fourth place, up 89 points; context window 1 million tokens, compared with 1,048,576 tokens for Argon.
  • Access first to a subset of cyber defenders through the Fairwind Program, then to trusted partners; no date given for developers, businesses or consumers.
  • Pricing: 2 dollars per million input tokens and 10 dollars per million output tokens introductory, rising to 4 dollars and 20 dollars after the introductory period, with 95% off cached input.
  • Google is taking part in United States government voluntary pre-release testing and safeguards are still being tested; benchmark figures are Google's own and independent verification is pending.

What Happens Next: Limited Release and Independent Testing

What happens next will depend on three things: how Carbon performs for the cyber defenders who get it first, whether independent testers can reproduce Google's benchmark results, and how rivals respond. The Fairwind Program will give Google feedback from real defenders working on real systems, which is a stronger test than a benchmark alone. If those defenders find the model useful and reliable, pressure will grow for wider access. If they find errors, false alarms or patches that break code, Google will need to improve the model or limit its use to supervised settings. Trusted partners will be the next group, and only after that stage is there likely to be a date for broader developer or business access.

Rivals are expected to respond soon, as they have after every major model launch in the past two years. New benchmarks, new models and new claims will follow, and the one-point lead on the Artificial Analysis index could change quickly. For readers, the practical advice is to treat Gemini 4 Carbon as a promising but unproven specialist model, not as a product they can use today. Do not assume it is publicly available, do not rely on benchmark scores alone, and watch for independent evaluations, pricing after the introductory period, and the results of government pre-release testing. Those steps will show whether Carbon lives up to its billing as Google's strongest model for cyber defence and long engineering work.

Sources: Reuters, Google DeepMind, October 2026.