Google Unveils Gemini 4 Carbon: A Cybersecurity-Focused Frontier Model with Restricted Access

Google has announced Gemini 4 Carbon, its latest frontier AI model, just two weeks after launching Gemini 4 Argon. Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, outperforming rivals like Claude Opus 5.5 and GPT-6 Astra. The model emphasizes cybersecurity, achieving state-of-the-art results in vulnerability detection and secure code generation. However, public access is restricted; Google is rolling it out first to vetted cyber defenders via the Fairwind Program, with no confirmed timeline for general availability.
Key points
- Gemini 4 Carbon is positioned as an incremental improvement over Gemini 4 Argon, with marginal gains in coding, agentic tasks, and scientific analysis.
- The model leads the Artificial Analysis Intelligence Index v5.0 with a score of 59, narrowly ahead of Claude Opus 5.5.
- Carbon excels in cybersecurity benchmarks, including 88.2% on SEC-bench Pro for vulnerability detection and 71.0% on CWE-bench for patching.
- Google is adopting a phased rollout, initially limiting access to cybersecurity defenders through the Fairwind Program and participating in U.S. government pre-release testing.
- Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens, with rates doubling after the promotional period.
Background
Google has been accelerating its release of frontier AI models to compete with OpenAI and Anthropic. In late September 2026, the company launched Gemini 4 Argon, which was initially restricted to cybersecurity partners before expanding to Pro and Ultra subscribers in mid-October. Internal testing of a model named 'Carbon' was reported in early October, suggesting it was a post-training upgrade of Argon. The rapid succession of releases reflects the intense competition in the AI frontier, where each lab claims a new lead within weeks of the previous release.
How outlets are covering it
Ars Technica characterizes Gemini 4 Carbon as a 'slightly better version of Argon,' noting that it outperforms competitors by slim margins and is not yet available to the public. The outlet questions whether Carbon is a new model or a post-training upgrade, given its similarities to Argon. Google's official blog post, authored by Koray Kavukcuoglu, emphasizes Carbon's state-of-the-art cybersecurity capabilities and long-horizon task performance, highlighting its top ranking on the Artificial Analysis Intelligence Index. Mashable focuses on the competitive dynamics, noting that Carbon's one-point lead over Claude Opus 5.5 is thin and could be erased by future releases. It also highlights the phased rollout strategy, which allows Google to claim leadership on leaderboards before public access. Futu (via WallStreetCN) provides a financial perspective, noting that Carbon's introductory pricing is half the standard rate, and that its cost per task is 23% lower than Claude Opus 5.5, though higher than GPT-6 Astra. The outlet also raises concerns about the dual-use nature of cybersecurity capabilities, emphasizing the need for robust safeguards.
Why it matters
The release of Gemini 4 Carbon underscores the rapid pace of innovation in frontier AI models and the increasing focus on cybersecurity capabilities. As AI models become more capable in vulnerability detection and code generation, the balance between expanding commercial applications and mitigating misuse becomes critical. Google's phased rollout strategy, involving government pre-release testing and limited access to cyber defenders, reflects a cautious approach to deploying powerful AI tools. The model's performance on independent benchmarks, such as the Artificial Analysis Intelligence Index, provides a benchmark for evaluating the competitive landscape, though the thin margins between models suggest that leadership is fluid and subject to rapid change.
What to watch
Google will continue to refine Gemini 4 Carbon's safeguards based on feedback from the initial cohort of cyber defenders in the Fairwind Program. The company plans to gradually expand access to developers, businesses, and consumers, though no specific timeline has been provided. If the rollout follows the pattern of Gemini 4 Argon, general availability could occur within a couple of weeks. The introductory API pricing will remain in effect until an unspecified date, after which rates will double to $4 per million input tokens and $20 per million output tokens. Competitors, including OpenAI and Anthropic, are expected to respond with new model releases, potentially erasing Carbon's current lead on leaderboards.
- Google announces Gemini 4 Carbon, its latest frontier AI model that you can’t use Ars Technica
- Gemini 4 Carbon: our next frontier model blog.google
- Google announces Gemini 4 Carbon, its new frontier AI model Mashable
- Google's cybersecurity powerhouse, Gemini 4 Carbon, has topped the AI intelligence rankings, scoring 88.2% in vulnerability detection, with a list price on par with Argon. 富途牛牛
- Google debuts flagship Gemini 4 Carbon model with cybersecurity upgrades SiliconANGLE
Want the full story? Read the original reporting
Read on Ars Technica