Google released its newest frontier model on Wednesday. Gemini 4 Carbon is the company’s latest attempt to lead the AI race on security and coding tasks.
The model posts “significant gains” in coding, agentic work, and cybersecurity, according to The Verge. Google also says Carbon now sits at the top of the Artificial Analysis Intelligence Index, version 5.0.
That ranking places it ahead of Claude Opus 5.5 and GPT-6 Astra. Both rivals had held strong positions on the index through the autumn.
Built to hunt for real bugs
Carbon arrives as an upgrade to Argon, which Google shipped earlier this month. Access is narrow for now and limited to a select group of cybersecurity defense companies.
Google plans to widen availability to additional partners through its Fairwind Program. The company has not given a firm date for that expansion or named the partners involved.
The security claims rest on two benchmarks. Carbon holds the number one spot on SEC-bench Pro, which measures a model’s ability to discover real-world bugs. It also leads ExploitGym, which tests how well a system turns a vulnerability into a working exploit.
Google says Carbon is built for longer tasks. Internal teams have watched it “run for days to persist through dead ends” on complex, multi-step engineering work.
Prompt injection and honesty
The company also claims better resistance to prompt injection attacks. That matters for defense work, where an outside party can try to steer a model mid-task.
Carbon is meant to avoid shortcuts and collaborate more closely with users. Google says it cuts “unsupported claims” by two times compared with Argon.
Pricing stays flat against the previous model. Carbon launches at $2 per million input tokens and $10 per million output tokens.
The release lands in a crowded month for frontier models. Google’s own Genie 4 world model shipped this week, and Anthropic’s IPO filing put fresh attention on Claude’s commercial future.
For security teams, the practical question is access. Most defenders cannot test Carbon yet, and the benchmarks Google cites measure lab performance rather than live incident response.
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
