Google launches Gemini 4 Carbon with focus on cyberdefense and coding

Google has released Gemini 4 Carbon, its newest frontier model, with claimed gains in coding, agentic workflows and cybersecurity. The launch matters for developers and security teams because Google is positioning Carbon not just as a general assistant, but as a model built to work through long engineering and cyberdefense tasks.

For now, access is narrow. According to Google, the model is available only to a select group of cybersecurity defense companies, with expansion planned later for other partners in its Fairwind Program.

What Google says is new in Carbon

Gemini 4 Carbon is the follow-up to Gemini 4 Argon, which Google released earlier this month. The company says Carbon improves on Argon across coding, agentic tasks and cybersecurity, and frames the release as evidence of how quickly it is moving its frontier models forward.

Google Gemini is already used as a general AI assistant and developer-facing model family, but Carbon’s launch is more specialized. Google is emphasizing secure code generation, vulnerability discovery and the ability to persist through complicated engineering work rather than only short prompt-and-response use cases.

Google says its internal teams have seen Carbon run for days while working through dead ends on complex, multi-step engineering tasks. That claim is important for agent workflows, where the practical limitation is often not whether a model can answer one question, but whether it can keep making useful progress without taking shortcuts or drifting away from the objective.

The company also says Carbon is better at resisting prompt injection attacks, avoiding shortcuts and collaborating with users. Compared with Argon, Google says Carbon reduces unsupported claims by two times, which it describes as more transparency and honesty.

Benchmark claims put security in the spotlight

Google says Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra. The Verge reports that Carbon is also ahead of those models on some benchmarks, based on Google’s announcement and benchmark materials.

The security-specific results are central to Google’s pitch. According to the company, Carbon currently holds the top position on SEC-bench Pro, which evaluates a model’s ability to find real-world bugs. Google also says it leads on ExploitGym, a benchmark focused on whether a model can turn a vulnerability into exploits.

Those results will be most relevant to teams already testing AI for secure development, red-team support, code review or vulnerability triage. For small software teams and freelance developers, the immediate takeaway is more indirect: frontier model makers are competing not only on chat quality, but on whether their systems can handle security-sensitive engineering work with fewer false claims and better resistance to hostile inputs.

Access is limited, and that matters

The early rollout is not a broad public launch. Google says Gemini 4 Carbon is currently available only to selected cybersecurity defense companies, before later expansion to additional Fairwind Program partners.

That limited access changes how most builders should evaluate the news. If you are choosing tools today for app development, code assistance or workflow automation, Carbon may not be something you can immediately plug into your stack. Existing coding assistants and agent tools remain the practical options for many teams while Google tests Carbon with security-focused organizations.

The pricing Google disclosed matches Argon’s competitive introductory API pricing: $2 per million input tokens and $10 per million output tokens. Google did not describe broader public availability in the source report, so teams outside the initial group should treat Carbon as a model to watch rather than a tool to budget around immediately.

For context, Carbon follows Google’s earlier Gemini 4 Argon release, which we covered when Google announced its cheaper introductory API pricing. It also fits a wider push toward safer AI agents, including Google’s work on isolated agent infrastructure such as Agent Compute Engine.

What creators and small teams should watch next

Carbon’s first audience is cybersecurity defense companies, not creators, marketers or general small businesses. Still, the direction of travel matters for anyone using AI to ship software, manage websites or automate client work.

If Google’s claims hold up in broader use, models like Carbon could make AI-assisted development more useful in areas where mistakes are costly: authentication flows, payment code, customer data handling and infrastructure scripts. But the same source report also highlights why access and guardrails matter: Carbon is evaluated not only on finding vulnerabilities, but also on exploit-related tasks.

For now, the practical checklist is simple:

• Security teams should track whether Carbon becomes available through the Fairwind Program.
• Developers should compare Google’s benchmark claims with hands-on results when access expands.
• Small businesses should avoid assuming benchmark wins translate directly into safer production code without review.

The launch shows where frontier AI competition is moving next: longer-running agents, stronger coding support and more emphasis on cyberdefense. Carbon may not be widely reachable yet, but it signals that model vendors are treating security performance as a headline feature, not an afterthought.