Google launches Gemini 4 Carbon, a frontier model honed for cybersecurity

What Google announced

Google DeepMind's announcement on 28 October 2026 introduces Gemini 4 Carbon as 'our strongest model with state-of-the-art cybersecurity and long horizon capabilities.' It arrives two weeks after Gemini 4 Argon, signaling Google's fastest frontier-model cadence yet.Carbon ranks #1 on the v5.0 Artificial Analysis Intelligence Index at 59, with gains across coding and agentic tasks and a firm lead on cybersecurity defense benchmarks: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching.

Cybersecurity credentials

On internal benchmarks, Google says Carbon achieves more than a 2x uplift in success rate for generating secure code over Argon. The Chrome team reported Carbon performed more than 10x better than another lab's large frontier model when generating high-quality patches.

Wiz, the cloud security platform, observes Carbon leading on benchmarks for web penetration testing, a core capability for its Scan for Good initiative. Google also says Carbon resists prompt injection at 0.0% attack success on the Gray Swan Indirect Prompt Injection Benchmark, down from Argon's 0.7%, and reduced successful cheating on Terminal-Bench 4.0 to 0.0% with attempts to cheat dropping to 0.4%.

Long-horizon engineering

Beyond security, Google positions Carbon for sustained agentic workflows: 'Internally, our teams have seen Carbon run for days to persist through dead ends to tackle complex, multi-step engineering tasks like root-causing and patching subtle bugs across large codebases and low-level systems,' the announcement says.

Carbon builds on Argon's alignment work. Google says it trained the model to respond to complex or high-pressure tasks in the right way, further mitigated shortcut attempts in training, and observed high-severity misalignment alerts in live coding environments drop by almost 80%.

Pricing and rollout

Gemini 4 Carbon launches 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 the input price.

Rollout is deliberately phased: Google is actively engaged in the US government's voluntary process for pre-release model testing, and starting this week will make Carbon available to a subset of cyber defenders, gradually phasing access to additional trusted partners through the Fairwind Program. Reuters reported Carbon is being made available first to cybersecurity defenders as the company continues testing safeguards.

What it means

Carbon's framing is a direct challenge to OpenAI and Anthropic in two arenas at once: frontier coding and cybersecurity defense. The security scores are vendor-reported, and the '10x better than another lab's model' claim is anonymous; as with all such benchmarks, independent testing will matter.

The phased rollout to cyber defenders is notable on its own: Google is choosing to put its most security-capable model in the hands of defenders first, and the US government pre-release testing engagement reflects the sensitivity of models that can find and exploit vulnerabilities.

What to watch

Watch when Carbon reaches general availability and whether its security benchmark leads hold in independent testing.

Watch how the Fairwind Program expands beyond the initial subset of defenders, and what safeguards Google adds.

Watch whether the two-week Argon-to-Carbon cadence becomes the new normal for Gemini frontier releases.

The security benchmark context

The three benchmarks Google leads with Carbon each measure a different phase of the defensive cycle. SEC-bench Pro tests whether a model can find vulnerabilities in real codebases; ExploitGym tests whether it can turn those findings into working exploits; CWE-bench tests whether it can patch the root cause with a working code fix. Together they cover the full vulnerability lifecycle.

Those capabilities are why Google is being deliberate about rollout. A model that can both find and exploit vulnerabilities is dual-use in a serious way, and Google's engagement with the US government's voluntary pre-release testing process acknowledges that this category of model attracts scrutiny from both defenders and attackers.

How it compares with Argon

Carbon builds directly on Argon, which launched two weeks earlier. Google's positioning is that Carbon is the stronger model for long-horizon engineering and security work specifically, while Argon remains the foundation for the broader Gemini 4 family. The pace suggests Google is now comfortable shipping multiple frontier models quickly rather than one per year.

For developers, the practical differences are pricing parity with Argon and a model tuned for tasks where sustained agentic work and security matter most: root-causing bugs, patching across large codebases, pentesting and defensive cyber operations. Whether it beats Argon on general-purpose tasks is a separate question from the security and long-horizon focus the announcement emphasizes.

Caveats

All the headline numbers come from Google's own announcement. The 88.2% SEC-bench Pro, 54.8% ExploitGym, 71.0% CWE-bench, the 2x secure-code uplift and the anonymous 10x patch comparison are vendor-reported; independent labs will need to reproduce them.

The Artificial Analysis Intelligence Index #1 ranking is notable because it comes from an independent third party, but Intelligence Index v5 was itself recently updated, and index methodology changes make score-to-score comparisons across versions approximate.