Google Unveils Gemini 4 Carbon With Initial Access for Cybersecurity Defenders

Google announced Gemini 4 Carbon on October 28, 2026, expanding its AI model family two weeks after introducing Gemini 4 Argon. The announcement adds to competition with OpenAI and Anthropic in coding and cybersecurity.

The company positions Carbon as a model for complex engineering work, including tasks that require sustained, multistep execution.

Its cybersecurity capabilities cover finding vulnerabilities, testing whether attackers could exploit them and producing patches.

Google Reports Gains Across Security Benchmarks

Google published the following results from evaluations conducted without safety guardrails:

The company also reported more than double Argon’s success rate for generating secure code in internal tests.

These figures describe performance under particular evaluation conditions. They do not mean the model will find or fix the same proportion of weaknesses in every organisation’s software.

The distinction between identifying a possible flaw and demonstrating its impact also matters. A suspected vulnerability may require further investigation before a security team can judge its severity.

Likewise, a proposed patch still needs testing to establish whether it addresses the problem without disrupting legitimate functions.

Restricted Rollout Reflects Potential Misuse Risks

The same capabilities that help defenders investigate weaknesses can also assist attackers. That creates a deployment challenge: expanding access to useful security tools while limiting opportunities for misuse.

Google said it would begin with selected cyber defenders and gradually extend access to trusted partners through its Fairwind Program.

It also confirmed participation in the US government’s voluntary prerelease testing process. Its announcement gave no date for general availability.

Restricted access and benchmark performance address different questions. Access controls determine who can use a system and under what conditions; evaluations measure how it performs on selected tasks.

Neither alone establishes how reliably a model will operate across unfamiliar production environments.

Introductory Pricing Starts at $2 per Million Input Tokens

Google announced these token rates:

Cached inputs receive a 95% discount against the input-token rate. Google did not specify when the introductory period would end.

Tokens are units a model processes when receiving information and producing responses. Token rates therefore describe usage charges rather than a fixed monthly subscription.

The cost of completing a security investigation will depend on how much information the model processes and how much work the task requires.