Hardware Ready Ready to Bench?
Google has unveiled Gemini 4 Carbon, a new artificial intelligence model that the company describes as its most powerful ever created. The announcement comes two weeks after the introduction of Gemini 4 Argon and focuses specifically on cybersecurity and the execution of complex, long-running tasks, which can last for several days.
Given its ability to identify software vulnerabilities, Google has opted for a phased release. Currently, Carbon is accessible only to a select group of cybersecurity defense experts, while developers, businesses, and users must wait longer.
Gemini 4 Carbon: why it matters
According to reports from Google, Gemini 4 Carbon achieves significant results in identifying, verifying, and fixing vulnerabilities. In tests conducted by the company without active protections enabled, the model identified 88.2% of real vulnerabilities on SEC-bench Pro, 54.8% on ExploitGym (dedicated to exploiting flaws), and 71.0% on CWE-bench, which evaluates the ability to fix problems at their root. In secure code generation, internal tests indicate a success rate higher than that achieved with Argon.
The Chrome team reported that Carbon produced fixes more than 10 times better than those from a large AI model developed by another laboratory, while security firm Wiz ranked it first in its penetration testing of web applications.
Google highlights improvements to the model in areas such as programming, agents, scientific tasks, web development, and understanding images and other content. Carbon ranked first on the Artificial Analysis Intelligence Index v5.0 with 59 points and rose to fourth place on WebDev Arena with 1,768 points, gaining 89 points compared to its previous evaluation.
Within Google, the model has been used for days on complex engineering tasks, such as identifying and fixing bugs in large codebases, overcoming obstacles without interruption.
What changes and what are the effects
Carbon inherits the safety system of Gemini 4 Argon, presented in mid-October and already available in Italy, featuring real-time defenses against abuse, prompt injection protections, and reasoning monitoring. Google states that the new model produces about half as many unsupported claims as Argon and, in the Gray Swan benchmark, no indirect prompt injection attacks succeeded, unlike the 0.7% recorded with the previous model. The most severe alarms regarding anomalous behaviors in its internal programming environments have decreased by nearly 80%.
Google describes a cautious approach to release while gathering feedback to improve protections. Starting this week, Carbon is available to a selected group of cybersecurity defense experts and will subsequently be extended to partners of the Fairwind program. Google also participates in voluntary tests conducted by the U.S. government on models before their publication and promises to make it available to developers, businesses, and users “as soon as possible,” without specifying a date.
Pricing for APIs has already been defined: during the launch period, prices will remain those of Argon, namely $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached tokens. At the end of the promotional period, prices will rise to $4 and $20 respectively.
Source and further reading on Gemini 4 Carbon: original article.
* Content created with the assistance of artificial intelligence systems.