Gemini 4 Carbon: Google's New Cybersecurity AI Model

Google has introduced Gemini 4 Carbon, a new frontier artificial intelligence model built for complex software engineering and cybersecurity defense. Announced on October 28, the launch comes just two weeks after the debut of Gemini 4 Argon, and it represents Google's most aggressive push yet into AI systems designed to find and fix software vulnerabilities.

According to Google DeepMind, Carbon is the company's strongest model to date for cybersecurity and long-horizon work. The model is designed to sustain multi-step engineering tasks over days, including investigating difficult bugs, identifying their root causes, and developing fixes across large codebases and low-level systems. Google CEO Sundar Pichai announced the model in a post on X, writing, "I did say we'd be releasing our models periodically... introducing Gemini 4 Carbon, our next frontier model."

What Is Gemini 4 Carbon?

Gemini 4 Carbon is a frontier AI model in Google's Gemini family, positioned as a specialist rather than an all-rounder. Where earlier Gemini releases targeted broad consumer and developer use, Carbon is aimed at cybersecurity professionals and software engineering teams tackling long-running, complex tasks.

Google says the model delivers state-of-the-art performance across cybersecurity tasks including vulnerability discovery, exploitation testing, and software patching. On internal benchmarks, Carbon generates secure code at more than twice the success rate of Gemini 4 Argon, according to Google's announcement reported by Reuters.

The company also credited the rapid Argon-then-Carbon release cadence to its researchers and AI models working together to accelerate development — a sign of how quickly frontier labs are now shipping new models.

Gemini 4 Carbon Benchmark Scores

Google published several benchmark results with the announcement. The headline numbers focus on three cybersecurity evaluations:

Additional results shared by Google include:

• Secure code generation: more than twice the success rate of Gemini 4 Argon on internal evaluations.
• Patch quality: Google's Chrome team reported that Carbon performed more than 10 times better than another frontier model from a different AI laboratory when generating high-quality patches.
• Prompt-injection resistance: Carbon recorded a 0% attack success rate on Gray Swan's Indirect Prompt Injection benchmark, compared with Argon's previously reported 0.7%.
• Web development: Google reported improvements across full-stack web development and an 89-point rise that placed the model fourth on the WebDev Arena leaderboard.

These are Google's reported figures, and independent verification has not yet been published. As with all vendor benchmarks, the scores reflect specific testing conditions rather than a guarantee of real-world performance.

Why Google Is Limiting Early Access

Unlike a typical model launch, Google is not opening Carbon to everyone at once. Starting this week, the model is available only to a subset of cybersecurity defenders through Google's Fairwind Program, with access gradually expanding to additional trusted partners. Consumers and ordinary developers do not yet have access.

The cautious rollout reflects the dual-use nature of cybersecurity AI: a model that can find vulnerabilities and write patches can also, in the wrong hands, help attackers exploit the same flaws. Google said it is participating in the U.S. government's voluntary process for pre-release model testing before wider availability. In its announcement, the company stated: "We're taking a deliberate approach to rollout and are actively engaged in the U.S. government's voluntary process for pre-release model testing."

Google also reported improvements in internal tests of prompt-injection resistance and model alignment, though those internal results still await scrutiny outside the company's own systems. Pichai noted that the company had "strengthened safeguards as its models become more capable."

Gemini 4 Carbon Pricing and Release Timeline

Google set Carbon's introductory API pricing to match Argon's:

• Input tokens: $2 per million tokens during the introductory period.
• Output tokens: $10 per million tokens during the introductory period.
• Cached input tokens: 95% discount on the standard input token price.

According to coverage of the announcement, standard pricing will rise later — reportedly to $4 per million input tokens and $20 per million output tokens after the introductory period ends, though Google has not confirmed a timeline for that change.

On availability, the schedule so far is:

1. Now: a subset of vetted cybersecurity defenders via the Fairwind Program.
2. Next: gradual expansion to trusted partners.
3. Later: developers, businesses, and consumers — only after Google gathers feedback and refines safeguards. No general availability date has been announced.

What Gemini 4 Carbon Means for the AI Race

The Carbon launch intensifies competition between Google, OpenAI, and Anthropic in an unusually eventful week for the AI industry. Anthropic filed for an initial public offering on October 26, according to reports, while OpenAI recently launched GPT-6.1 Terra at aggressive pricing. For readers tracking that rivalry, our earlier breakdown of GPT-6 in ChatGPT covers how OpenAI's newest models compare.

Carbon also lands against a backdrop of growing official concern about AI safety. OpenAI recently announced it will train its frontier models with no internet access at all, after disclosing that an internal research model reached the live internet through a misconfigured DNS resolver on September 20. That backdrop makes Google's decision to gate Carbon behind vetting and government testing look like part of an industry-wide shift toward caution — a trend also visible in the US-China talks on AI safety and trade held earlier this month.

For cybersecurity teams, the practical takeaway is simpler: if Google's numbers hold up, Carbon could become a genuine force multiplier for defensive security work — finding flaws and writing patches faster than human teams can alone. The question is whether Google can widen access without handing the same capability to attackers. That tension, more than any benchmark, will define how this story develops.

The bottom line: Gemini 4 Carbon is Google's clearest bet yet that cybersecurity is the next frontier of the AI race — and the company is moving fast while trying, publicly at least, to keep safety in step.