Google has officially introduced Gemini 4 Carbon, an advanced artificial intelligence model specifically engineered to tackle demanding coding challenges and assist cybersecurity professionals in detecting and remediating software vulnerabilities. Unlike general-purpose chatbots, Gemini 4 Carbon is built for multi-step technical problem-solving that requires sustained effort.
Limited Initial Access for Cybersecurity Defenders
Currently, access to Gemini 4 Carbon is not broadly available. Google is implementing a cautious, phased deployment, initially offering the model to a select group of cybersecurity defenders. The company plans to gradually expand access to other trusted partners in the future, with broader availability for developers, businesses, and consumers anticipated at an unannounced later date.
Key Capabilities and Performance
One of the primary functions of Gemini 4 Carbon is its ability to pinpoint weaknesses in software that could be exploited by malicious actors. It can also help security teams evaluate the exploitability of a vulnerability and generate code modifications to resolve the underlying issue. For example, if a developer faces a complex bug in a large software project, Carbon is designed to investigate its root cause and assist in developing a fix. Google reports that its internal teams have observed the model working for days on intricate engineering tasks, including problems requiring multiple steps and iterative attempts at solutions.
Strong Performance on Benchmarks
- SEC-bench Pro: Gemini 4 Carbon scored 88.2% in identifying software vulnerabilities.
- ExploitGym: It achieved 54.8% in evaluating vulnerability exploitation.
- CWE-bench: The model scored 71% in assessing the ability to fix vulnerabilities.
Google notes these benchmark results were achieved without guardrails and do not guarantee identical real-world performance. In internal testing, Carbon also reportedly more than doubled the success rate of secure-code generation compared to its predecessor, Gemini 4 Argon.
Pricing and Strategic Deployment
The decision to limit initial access stems from Google's recognition that while AI models capable of identifying and testing software vulnerabilities are powerful tools for protection, these same capabilities could potentially be misused. The gradual deployment allows Google to continue evaluating safeguards and gather crucial feedback from early adopters.
For developers accessing the model via an application programming interface (API), Google has announced introductory pricing:
- Input Tokens: $2 per million tokens
- Output Tokens: $10 per million tokens
- Cached Input Tokens: 95% discount on input-token rate
After the introductory period, these prices are set to increase to $4 per million input tokens and $20 per million output tokens. Gemini 4 Carbon is currently targeted at cybersecurity professionals and developers engaged in demanding technical work, rather than general-purpose AI users.