Google introduces Gemini 4 Carbon for coding and cybersecurity
Google has announced Gemini 4 Carbon, a model designed for complex coding and cyber defence tasks. Access will initially be limited to a group of cybersecurity professionals, with no date announced for general availability.
Summary
Google has introduced Gemini 4 Carbon, an AI model designed for long-running software engineering and cybersecurity tasks. According to the company, the model can help identify vulnerabilities, assess whether they can be exploited and produce fixes for the problems found.
The announcement comes two weeks after Gemini 4 Argon. Google says Carbon improves performance in coding, reasoning and agentic tasks, but initial access will be limited to a group of cyber defenders.
In practice
In internal tests, Google says it has observed Carbon working for days on multi-step engineering problems, including finding the root cause of subtle bugs and patching them across large codebases.
For cybersecurity, the company reports scores of 88.2% on SEC-bench Pro for vulnerability detection, 54.8% on ExploitGym for exploitation and 71% on CWE-bench for patching. Google disclosed these results and says the tests were conducted without usage safeguards enabled, so they should not be treated as a guarantee of performance in products or real-world environments.
Google also announced introductory pricing of $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached input tokens.
What we still don't know
Access to Carbon will begin with a limited group of cyber defenders and expand gradually to trusted partners through the Fairwind Program. Google intends to make it available to developers, businesses and consumers later, but has not announced a date.
The company says it is still gathering feedback and refining safeguards before expanding access. It also says it is participating in the US government’s voluntary process for pre-release model testing.
In the independent Artificial Analysis index, the Carbon High configuration scores 59 points, placing it first among the models evaluated. The ranking combines several assessments, but does not replace testing against an organisation’s own tasks and conditions.
Why it matters
- The model points towards longer-running use of AI agents on multi-step software problems.
- Finding, testing and patching vulnerabilities could support security teams, although results will depend on how the model is used.
- Restricted access reflects the need to assess safeguards for cybersecurity capabilities that could also be misused.