Google released Gemini 4 Carbon on Wednesday, its latest frontier model. Additionally, the company says the model brings a “significant leap” in cybersecurity, according to its announcement. Indeed, the model tops the Artificial Analysis Intelligence Index (v5.0) with 59 points. Claude Opus 5.5 and GPT-6 Astra trail behind. In fact, Google calls Carbon its strongest model for long-horizon engineering and cyber defense.
Gemini 4 Carbon arrives just two weeks after Gemini 4 Argon. That pace is the point, Google argues. Researchers and models now advance capabilities together. Also, the new model is available first to a select group of cyber defenders. As a result, wider access will follow through Google’s Fairwind Program. Developers, businesses, and consumers will get it later, after more feedback and guardrail work.
Meanwhile, the pricing stays flat. At launch, Carbon costs the same as Argon. Input tokens are $2 per million. Output tokens are $10 per million. Cached input tokens get a 95 percent discount. After the introductory period, those rates rise to $4 and $20 per million respectively.
What Gemini 4 Carbon can do
Furthermore, the model is built for sustained agentic work. Internal teams ran it for days. Notably, the model persisted through dead ends and fixed subtle bugs across large codebases. Google also reports gains in science tasks, multimodal understanding, and full-stack web work.
On security benchmarks, Carbon leads. For example, it scores 88.2 percent on SEC-bench Pro, a test of autonomous flaw discovery in real codebases. Specifically, the model reaches 54.8 percent on ExploitGym, which tests turning vulnerabilities into working exploits. Likewise, the model hits 71.0 percent on CWE-bench for writing fixes. These tests ran without safety guardrails.
Notably, Carbon also resists prompt injection. On Gray Swan’s indirect prompt injection benchmark, the attack success rate fell. In other words, the score dropped from 0.7 percent with Argon to 0.0 percent. Still, that does not solve prompt injection. The test may be saturated.
Independent results back some of the claims. Artificial Analysis places Carbon first on its Intelligence Index v5.0. The model scores 70.36 percent on the Vals AI index, second behind Claude Sonnet 5.5. On Arena, Gemini 4 Carbon (High) leads the Text Arena with 1,528 points. Then, the model places second in the Agent Arena. The model’s median task cost is $1.80, far below Claude Opus 5.5 (Max).
In addition, Google reports internal gains that outsiders cannot verify. Carbon produces secure code at more than double Argon’s rate. Moreover, the Chrome team says it performed more than tenfold better than a rival lab’s top model when writing patches. Wiz, a cloud security platform, sees Carbon leading on web penetration testing benchmarks.
Why Google limited early access
Google is taking a deliberate approach. For Gemini 4 Carbon, that balance matters. Carbon is undergoing the U.S. government’s voluntary pre-release safety review. In fact, the reason is its cyber capability. A model that finds and exploits vulnerabilities can also help attackers. Google wants defenders first.
Meanwhile, safety work runs in parallel. On Terminal-Bench 4.0, successful cheating fell to zero. High-severity misalignment alerts dropped by almost 80 percent in internal coding environments. Unsupported claims fell by about 2 times compared with Argon. Overreach dropped 26 percent and under-service 29 percent.
However, there is a trade-off. Answer accuracy rose from 48 percent to 58 percent. However, the hallucination rate climbed from 4 percent to 11 percent. That is still the second-lowest rate among the models tested. Claude Opus 5.5 (Max), along with GPT-6 Astra (Max), both sit at 50 percent.
What to watch next for Gemini 4 Carbon
Ultimately, three questions matter now. First, when does broader access arrive? Google has not set a date for developers, businesses, or consumers. Second, what happens to pricing after the introductory discount ends? Third, how do rivals respond? Anthropic and OpenAI are also pushing agent and security features. OpenAI recently tapped NASK to stress-test frontier AI security, a sign that labs now treat model safety as a product.
Overall, the broader pattern is clear. Frontier models now ship every two weeks. The race has shifted from chat quality to long-running agents and cyber defense. Gemini 4 Carbon is Google’s bet. Gemini 4 Carbon shows that the next buyer of frontier AI will be a security team, not a chatbot user.
I am a person who is positive about every aspect of life.I have always been an achiever be it academics or professional life. I believe in success through hard work & dedication.
Technology Blogger at TechnoSecrets.com
