Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity
Developing story first seen 1 hour ago
Google has launched Gemini 4 Carbon, its latest frontier model, with claimed improvements in coding, agentic work and cybersecurity. The new development is its restricted initial release to selected cybersecurity defence companies, with wider access planned through Google’s Fairwind Program.
Google says Carbon ranks first on the Artificial Analysis Intelligence Index, SEC-bench Pro and ExploitGym. It is designed to handle longer engineering tasks, resist prompt injection and make half as many unsupported claims as its predecessor, Argon. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens.
- Sources
- 2
- Outlets
- 2
- Left · Centre · Right
- 0 · 0 · 0
In short
- Google has launched Gemini 4 Carbon.
- Selected cybersecurity firms get access first.
- It matches Argon’s introductory pricing.
How this story developed
- 44 minutes ago Google has released Gemini 4 Carbon, a new frontier AI model that it says makes significant advances in coding, multi-step tasks and cybersecurity.
- 50 minutes ago Google has launched Gemini 4 Carbon, its latest frontier model, with claimed improvements in coding, agentic work and cybersecurity. The new development is its restricted initial release to selected cybersecurity defence companies, with wider access planned through Google’s Fairwind Program.
Full account
Google has announced Gemini 4 Carbon, a new frontier AI model that it says improves on Gemini 4 Argon in coding, multi-step tasks and cybersecurity. The release comes about two weeks after Argon. Google describes Carbon as a substantial advance, while one report characterises the differences between the two models as modest and notes that the company has not explained whether Carbon is a wholly new model or an updated version of Argon.
The company highlights Carbon’s results on several benchmarks. It says the model ranks first on the Artificial Analysis Intelligence Index and on SEC-bench Pro, which assesses the discovery of real-world software bugs, as well as ExploitGym, which tests whether models can develop exploits from vulnerabilities. Google also reports an 89-point rise on WebDev Arena, placing Carbon fourth. These are company-cited comparisons; benchmark scores offer one measure of performance rather than a complete account of how a model will fare in use.
Google says Carbon is better at identifying and addressing vulnerabilities, producing secure code and persisting through long engineering tasks. Its internal teams have reportedly used it on work lasting days. The company also says Carbon is less prone to unsupported claims than Argon. Separately, the Chrome team says it produces high-quality patches at a higher rate than non-Google models. On prompt-injection testing by Grey Swan, Google reports that Carbon resisted all attempts, compared with small failure rates for Argon and Claude Opus 5.5 and a higher rate for GPT-6 Astra.
Access is restricted for now. Google says it will first make Carbon available to selected cybersecurity defence companies, then extend access through its Fairwind programme after a deliberative testing process. The reports describe a staged release but give no firm date for wider availability. One suggests developers using the Gemini API and AI Ultra subscribers may be among the first later users. Google says the introductory API price will match Argon’s: $2 per million input tokens and $10 per million output tokens.
Where outlets differ
From the full account, which is AI-written from the sources:
The first report foregrounds Carbon’s headline benchmark positions, its limited availability to cybersecurity companies and Google’s stated price. The second is more sceptical about the scale and significance of the improvements, observing that Carbon’s margins over competitors are small and questioning whether it is a new model or a post-training update to Argon.
The second report adds claims about vulnerability-free code, Chrome patch generation, WebDev Arena and Grey Swan prompt-injection results that are not included in the first. It also speculates about possible early access for API developers and AI Ultra subscribers; Google has not provided a firm timeline for general availability.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Restricting initial access to vetted cybersecurity defenders could give them advanced tools to find and fix vulnerabilities while limiting opportunities for criminals to exploit the same capabilities. A staged rollout through the Fairwind Program may also help Google assess risks and improve safeguards before broader release.
The case against
Keeping a powerful model from wider access may concentrate cybersecurity advantages among a small group of selected companies, while independent researchers and smaller organisations lack comparable tools. Critics could also question whether company-reported benchmark results and promised safeguards are enough to justify restricted access, especially if the selection process is not transparent.
Explained simply
Google develops the Gemini family of artificial intelligence models, including systems that can write code, analyse information and carry out multi-step tasks. Its advanced models are often called “frontier” models, meaning they are among the most capable systems available. Cybersecurity teams may use them to find weaknesses in software and help defend computer networks.
The main players are Google and the security companies that build tools to protect computer systems. More capable AI could help defenders spot problems faster, but the same abilities could also be used to find or exploit weaknesses. How access is managed matters because it can affect who can use these tools and for what purposes.
Coverage · 2 sources
Americas Art Climate Culture Cybersecurity Science Technology World