Google’s Carbon Targets Cybersecurity With Longer, Harder AI Tasks

Google announced Gemini 4 Carbon on October 28, describing it as “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” The model focuses on cybersecurity, complex coding and tasks that may take days to complete, including tracing bugs across large codebases and low-level systems.

Carbon builds on Gemini 4 Argon, which launched two weeks earlier. Google said the new model improves vulnerability detection, security-risk assessment and code fixes, while also advancing coding, scientific work, web development and AI agents that carry out multi-step workflows.

The model will not receive a broad public release at launch. Google plans to start with selected cybersecurity defenders, then expand access to additional trusted partners through its Fairwind Program while it continues pre-release testing and work on safeguards.

A model built for vulnerabilities and complex fixes

Google said Carbon can identify and prioritize software vulnerabilities, validate those risks and patch them at their root cause. That focus sets the model apart from a system that only points out a flaw, because Carbon is designed to help develop the code fixes needed to address security problems.

On SEC-bench Pro, a benchmark for identifying vulnerabilities in real-world software, Carbon scored 88.2%. It scored 54.8% on ExploitGym, which evaluates vulnerability exploitation, and 71.0% on CWE-bench, which measures the ability to fix software weaknesses.

Google said these external cybersecurity evaluations took place without guardrails. The company also said internal benchmarks showed Carbon generated secure code at more than twice Argon’s success rate. Google’s Chrome team reported that Carbon performed more than 10 times better than another frontier AI model from a different lab when generating “high-quality” patches.

Google did not identify that competing model or explain the comparison method in its announcement. The company also said Carbon reduces shortcuts and loophole exploitation, improves transparency and honesty, resists prompt injection, and follows through on tasks without overstepping.

Strong benchmark results, careful rollout

Artificial Analysis ranked Carbon first on its Intelligence Index version 5.0, giving it a score of 59. Google said the model tops Claude Opus 5.5 in several key benchmarks and posted leading results across agentic coding, computer use, knowledge work, science and math, along with other categories.

Carbon also improved its WebDev Arena score by 89 points. It ranked fourth on that leaderboard with a score of 1,769, giving Google another measure of the model’s coding and web-development performance.

In the Grey Swan evaluation, Carbon resisted every attempted prompt-injection attack, receiving a 0 percent score. Google listed Argon at 0.7 percent, Claude Opus 5.5 at 1 percent and GPT-6 Astra at 8.5 percent in the same comparison.

Those cybersecurity and agentic-coding results explain Google’s cautious release plan. The company said it is taking a “deliberate approach” and participating in the U.S. government’s voluntary process for pre-release model testing.

“Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program,” Google said. The company added that it will gather feedback and refine its guardrails before making Carbon available to developers, businesses and consumers through its products.

Pricing starts at Argon’s introductory rates

Gemini 4 Carbon’s introductory price matches Argon’s rates: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount against the standard input-token price.

After the introductory period, Google said pricing will rise to $4 per million input tokens and $20 per million output tokens. Google has not announced an end date for the introductory pricing or a specific date for public access.

The release plan mirrors Argon’s rollout. Carbon will first go to selected cybersecurity defense companies, followed by additional trusted partners through the Fairwind Program, before Google makes a wider release.

Google said it plans to make Carbon available to developers, businesses and consumers as soon as possible, subject to the phased rollout and continued safety work. AI Ultra subscribers and developers using the Gemini API may receive access first, but Google has not promised a timeline for that wider availability.

“Our researchers and models are working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem,” Google said. For Carbon, the central test will be whether its ability to handle long-running engineering work can expand without creating new security risks.

Stay connected via Google News

Follow us for the latest AI News