Google on Oct. 28 announced Gemini 4 Carbon, a new artificial intelligence model that the company called its strongest yet, with a focus on cybersecurity and long-running engineering tasks. Google is starting with a limited release to a subset of cyber defenders, people who work to protect computer systems from attack.
According to Google, the rollout is deliberate. The company said it is taking part in the U.S. government’s voluntary process for pre-release model testing. It said it will add more trusted partners gradually through its Fairwind Program and is gathering feedback and adjusting its guardrails before offering Carbon to developers, businesses and consumers “as soon as possible.”
Koray Kavukcuoglu, senior vice president of Google DeepMind and Chief AI Architect at Google, wrote the announcement. It said Carbon builds on Gemini 4 Argon, which Google launched two weeks earlier.
Google said Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. The company reported results on three outside cybersecurity tests, all evaluated without guardrails. It said Carbon scored 88.2% on SEC-bench Pro, which measures finding vulnerabilities, 54.8% on ExploitGym, which measures exploiting them, and 71.0% on CWE-bench, which measures patching them. Google described each as a state-of-the-art or leading result.
Google also made claims based on internal testing. It said Carbon more than doubled Argon’s success rate in generating secure code. It said the Chrome team reported that Carbon performed more than 10 times better than another lab’s large frontier model when generating high-quality patches.
Google also said Wiz, which it called its leading cloud security platform, observed Carbon leading on benchmarks for web penetration testing. That is testing a system’s defenses by simulating attacks.
On long tasks, Google said its teams saw Carbon run for days on multi-step engineering work, such as finding the root cause of subtle bugs in large codebases and fixing them. In web development, the company said Carbon gained 89 elo points, a ranking score, to place fourth on WebDev Arena with 1768.
Google said Carbon uses the same safety and security approach as Argon, which combines model training, real-time defenses against misuse, resistance to prompt injection and monitoring of the model’s reasoning. Prompt injection is an attack that hides malicious instructions in content a model reads. Google said Carbon had a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection Benchmark, down from 0.7% for Argon.
Google said that in internal testing, Carbon cut unsupported claims by about half compared with Argon. It said high-severity misalignment alerts in live coding with its own teams dropped by almost 80%. It also said successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%.
Google said Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Tokens are the small units of text that AI models process. Cached input tokens will cost 95% less than the input price. After the introductory period, Google said, the price will be $4 per million input tokens and $20 per million output tokens.