Google announces Gemini 4 Carbon with cybersecurity focus

Ars Technica reports that Google has announced Gemini 4 Carbon, its latest frontier AI model, just two weeks after the release of Gemini 4 Argon. The announcement highlights a strategic shift toward models with enhanced cybersecurity capabilities, specifically in vulnerability detection, validation, and remediation. Google states that Carbon is twice as likely to generate code free of vulnerabilities compared to Argon, positioning it as a significant upgrade for secure software development.

The release follows a pattern of staged rollouts, with Google citing the need for a "deliberative approach" due to the model's strong performance in agentic coding and cybersecurity. The company is working within the voluntary testing framework established by the US government and plans to roll the model out through its Fairwinds cyber defense program. While Google has not provided a specific timeline for wide availability, Ars Technica notes that if the release mirrors that of Argon, the model could be available in a couple of weeks.

Performance and Benchmarks

Google positions Carbon as a new model, though it shares many characteristics with Argon, potentially indicating a post-training upgrade. In head-to-head comparisons, Carbon outperforms Argon by slim margins in every test provided by Google. It also beats GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 in almost every benchmark. Notably, the Chrome team claims Carbon is 10 times better at generating "high-quality" patches compared to non-Google models.

On the WebDev Arena leaderboard, Carbon has increased its score by 89 points, currently ranking at number four. While these numbers place Carbon at the top of Google’s internal charts, Ars Technica points out that the gap between Carbon and competing frontier models is only a few points, suggesting that benchmarks alone do not tell the whole story. All frontier models score in the same neighborhood, with Carbon holding a slight edge.

Safety and Security

A key differentiator for Carbon is its performance in safety evaluations. In the Grey Swan evaluation, Carbon resisted every attempted prompt-injection attack, achieving a 0 percent failure rate. This score is superior to Argon (0.7 percent), Claude Opus 5.5 (1 percent), and GPT-6 Astra (8.5 percent). Google attributes this to more rigorous safeguards implemented in the model, reinforcing its positioning as a secure option for enterprise and developer use cases.

Pricing and Availability

For developers, the initial availability is expected to be limited to AI Ultra subscribers and those using the Gemini API. Google will offer a discounted API rate of $2 per million input tokens and $10 per million output tokens during an introductory period. This price is set to double once the introductory period expires. Unlike Argon, whose discounted rates are valid through the end of November, Google has not specified an end date for Carbon’s introductory pricing, declining to provide a timeline when asked.

What It Means for Developers

For developers working on security-critical applications, Carbon’s improved vulnerability detection and patch generation capabilities could reduce the manual effort required for code auditing. The 10x improvement in patch quality, as claimed by the Chrome team, suggests that automated security fixes may become more reliable. However, the staged release means that access will be limited initially, and the temporary pricing structure requires careful budget planning. Developers should monitor the Fairwinds cyber defense program updates for official release dates.

What to Watch

• Release Timeline: Watch for announcements from the Fairwinds cyber defense program, which will signal the start of the wide release.
• Pricing Changes: Keep an eye on the Gemini API pricing page for the expiration of the introductory $2/$10 per million token rates.
• Competitor Responses: Monitor how OpenAI and Anthropic respond to Carbon’s security benchmarks, particularly in prompt-injection resistance and patch generation.
• WebDev Arena Rankings: Track Carbon’s position on the WebDev Arena leaderboard as more models are evaluated.