Google launches Gemini 4 Carbon for cyber defenders first
Google's new Gemini 4 Carbon model dominates coding and security benchmarks, but only cyber defenders get access this week.
Models · Source: 9to5Google
What happened
Google just announced Gemini 4 Carbon. It comes only two weeks after their Argon release. The new frontier model focuses heavily on coding, agentic workflows, and cybersecurity defense. Researchers and models worked hand in hand to advance capabilities at a rapid pace. Carbon ranks number one on the recently updated version five of the Artificial Analysis Intelligence Index with a score of 59. It also jumps 89 points to land at number four on the WebDev Arena.
Carbon excels at finding and fixing security holes. It scored 88.2 percent on SEC-bench Pro for finding vulnerabilities. It hit 54.8 percent on ExploitGym for exploitation and 71 percent on CWE-bench for patching. Google says it cuts unsupported claims in half compared to Argon. The model stops itself from overreaching on multi-step plans by 26 percent. It also reduces under-service by 29 percent to ensure reliable execution.
You cannot use it yet. Google is restricting initial access. The model goes exclusively to cyber defenders this week. Developers, businesses, and consumers will get access later. Once available, Google will offer introductory pricing of two dollars per million input tokens and ten dollars per million output tokens. Cached input tokens will be priced at a massive 95 percent discount. Regular pricing will eventually settle at four dollars for input and twenty dollars for output.
Key facts
- 59 — Score on the Artificial Analysis Intelligence Index
- 88.2% — Score on SEC-bench Pro for finding vulnerabilities
- 71.0% — Score on CWE-bench for patching
- $2 — Introductory price per million input tokens
- 2x — Reduction in unsupported claims compared to Argon
Why it matters
Builders get a highly capable agentic coder that actually respects constraints. Carbon retrieves relevant sources instead of guessing when faced with ambiguity. It surfaces discrepancies to the user before acting. It puts methodological caveats directly into the deliverable. It withholds blocked actions with a clear list of what remains required. This means less time debugging AI hallucinations and more time shipping secure code. Live coding environments show reduced misalignment and fabrication rates.
Google is prioritizing defense over open access. By giving cybersecurity teams the tool first, they prevent attackers from exploiting the model's vulnerability-finding skills. Strong AI goes to defenders before it reaches the public. This staggered release strategy sets a new standard for deploying highly capable security models. It shows that as models get better at exploitation, companies must rethink how they launch them.
For builders
Cheaper cached inputs for large codebases
Google offers a 95 percent discount on cached input tokens. Founders can feed massive code repositories into Carbon without burning cash. This makes full-stack web development and deep security audits financially viable for small teams.
Automated security patching
Carbon scores high on exploitation and patching benchmarks. Engineering teams can use it to automatically find and fix vulnerabilities before deployment. Companies that pay for manual security audits will save money and move faster.
Better agentic task execution
Carbon reduces overreach and under-service in multi-step plans. Builders can trust the model to stay strictly within scope and respect project constraints. This lowers the risk of autonomous agents breaking production systems.
My take
Google built a model so good at finding security holes they had to lock it behind closed doors. Giving it to defenders first is a brilliant strategic move. If you build AI agents, take notes on how Carbon asks for help instead of hallucinating. Defense first, demo later.
Original reporting: 9to5Google. This is my rewrite and opinion.