Gemini 4 Carbon: Google DeepMind's Next Frontier Model

Google DeepMind has announced Gemini 4 Carbon, its next frontier model, and the company describes it as its strongest model yet, with state-of-the-art cybersecurity and long-horizon capabilities. The announcement, made on Oct. 28, 2026, matters because it points toward advances in coding, agentic tasks and defensive security, according to Google DeepMind.

What Google DeepMind announced

Google DeepMind said it is introducing Gemini 4 Carbon as its next frontier model. The company called it its strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon ranked first on the updated v5.0 of the Artificial Analysis Intelligence Index at 59, according to Google DeepMind, and showed significant gains across coding and agentic tasks as well as across key capabilities critical for cybersecurity defense. The source reported Carbon led on SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0%. It was built on Gemini 4 Argon, which launched two weeks earlier. Google DeepMind said Carbon would launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. After the introductory period, the price would be $4 per million input tokens and $20 per million output tokens.

How much progress does it show for coding and long-horizon tasks?

Google DeepMind said Carbon is built for much harder, long-horizon engineering problems and for sustained agentic workflows. The company said internal teams saw Carbon run for days and persist through dead ends to tackle complex, multi-step engineering tasks, such as root-causing and patching subtle bugs across large codebases and low-level systems. Google DeepMind said it focused on coding, agentic, scientific tasks, full-stack web development and multimodal understanding. With better quality of output and artifacts in web development, Carbon improved by 89 Elo to land at 1768 and rank fourth on WebDev Arena. On internal benchmarks, Carbon drove a strong uplift in generating secure code and, in live coding environments with Google DeepMind teams, showed reduced misalignment and fabrication rates.

The company also said Carbon illustrates the pace at which it is pushing the frontier forward. Researchers and models are working hand in hand to advance capabilities faster while strengthening safeguards in tandem. That combination is part of why the announcement matters for developers who have been watching long-horizon agentic coding systems.

What makes the cybersecurity performance notable?

The cybersecurity claims center on the defense lifecycle: finding vulnerabilities, validating risk and patching at the root cause. Google DeepMind said Carbon delivers state-of-the-art performance across that lifecycle. The company said Carbon achieved 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Google DeepMind said internal benchmarks showed more than 2x uplift in success rate for generating secure code over Argon. The Chrome team reportedly said Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches. Wiz, a cloud security platform, also observed Carbon leading on benchmarks for web penetration testing, a core capability for Wiz's Scan for Good initiative.

Google DeepMind said these were external benchmarks evaluated without guardrails. The results are significant because they suggest a frontier model moving from general coding benchmarks toward defensive security tasks, where vulnerability discovery, validation and remediation are tightly linked.

How is safety and rollout being handled?

Google DeepMind said Carbon builds on the same multi-layered safety and security stack of Argon, combining model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. The company said it made significant strides in model alignment with Argon and reported its propensities to behave in an aligned way in its Safety and Security report. With Carbon, Google DeepMind said it trained the model to respond to complex or high-pressure tasks in the right way.

Google DeepMind said Carbon reduced shortcut attempts in training while ensuring high-quality coding performance. It said Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4%, while improving overall accuracy. In live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80%. Google DeepMind said internal testing showed Carbon cut unsupported claims by approximately 2x over Argon and put methodological caveats into deliverables or withheld blocked actions with a clear list of what remained required.

The company also said Carbon improved collaboration with users in the face of ambiguity, reduced overreach by 26% and under-service by 29%, and resisted prompt injection at 0.0%, compared with Argon's 0.7%. On rollout, Google DeepMind said it was taking a deliberate approach and was actively engaged in the U.S. government's voluntary process for pre-release model testing. Starting that week, it would make Carbon available to a subset of cyber defenders and gradually phase access to additional trusted partners through the Fairwind Program. It said it would gather feedback and iterate on guardrails before making Carbon available to developers, businesses and consumers via product surfaces as soon as possible.

How could developers and creators use this in Grafy?

For coding-focused readers, the story touches on long-horizon coding and secure code generation. Grafy's Code app is an AI coding studio that installs packages and runs code, which is a relevant surface for testing such workflows, though Google DeepMind did not say Carbon is available through Grafy. The practical connection is to explore how frontier coding models handle extended tasks, secure code and agentic debugging. Readers who want to compare multiple models for a coding project can use Grafy Chat, and developers can route work through Grafy Code.

Frequently asked questions

What is Gemini 4 Carbon?

Google DeepMind announced Gemini 4 Carbon as its next frontier model, calling it its strongest model with cybersecurity and long horizon capabilities. Carbon reached number one on the v5.0 Artificial Analysis Intelligence Index at 59, according to the company, and led on several cybersecurity benchmarks. The source is at https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/.

Why is the cybersecurity performance significant?

The cybersecurity results were described by Google DeepMind as state of the art. Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. The company said internal tests showed more than two times uplift in secure code generation over Argon, and the Chrome team reported more than ten times better patch quality than another large frontier model. The gains matter for defensive security workflows.

When will developers get access?

Google DeepMind said the first rollout would start this week with a subset of cyber defenders, then gradually add trusted partners through the Fairwind Program. It said access for developers, businesses and consumers via product surfaces would come as soon as possible after feedback and guardrails. Carbon launched at the introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off.

---

Source: Google DeepMind — Gemini 4 Carbon: Our next frontier model