Situation report: Google has a new frontier model at the top of the leaderboard, and the first people allowed to use it are cyber defenders. Google released Gemini 4 Carbon on Wednesday, according to The Verge AI. The company says the model brings “significant gains” in coding, agentic tasks, and cybersecurity. Carbon comes only weeks after Argon, which Google shipped earlier this month.
This is a significant release for two reasons. First, it puts Google at the top of a major independent benchmark. Second, it shows how frontier labs now handle models that are very good at finding software flaws.
🎯 The Core Intel
As detailed in The Verge AI, here’s what Google announced:
- Top benchmark position. Carbon ranks first on the Artificial Analysis Intelligence Index (v5.0). Claude Opus 5.5 and GPT-6 Astra sit behind it.
- Security benchmarks. It also holds the top spot on SEC-bench Pro, which tests whether a model can find real-world bugs, and on ExploitGym, which tests whether it can turn a vulnerability into a working exploit.
- Limited access. For now, only a select group of cybersecurity defense companies can use it. Google plans to expand access to other partners in its Fairwind Program.
- Long-running tasks. Google says its internal teams have seen Carbon “run for days to persist through dead ends to tackle complex, multi-step engineering tasks.”
- Fewer unsupported claims. Google says Carbon cuts “unsupported claims” in half compared to Argon and is better at avoiding shortcuts, resisting prompt injection, and working with users.
- Same price as Argon. It launches at Argon’s introductory price: $2 per million input tokens and $10 per million output tokens.
🛡️ Why the Gated Rollout Matters
One capability sticks out from the rest. A model that tops ExploitGym is, almost by definition, a model that can help someone break into systems. A tool that finds bugs and writes exploits is useful to defenders and attackers alike.
Google is giving defenders a head start. Security firms get the model first, so they can find and patch vulnerabilities before something with similar skills becomes widely available. Over the past year this has become a familiar pattern across the industry. Labs increasingly see advanced cyber capability as something to release in stages, not all at once.
If you work in security, expect this to change your workflow. Automated bug hunting and secure code generation are moving from research demos into production tools. Your adversaries will get comparable tools eventually, too.
⚙️ The Agentic Angle
Google’s claim that Carbon can run “for days” deserves attention. Until recently, most agentic coding sessions lasted minutes or hours before the model lost the thread, looped, or quietly cut corners.
Google is targeting exactly those weaknesses. Two improvements matter most:
- Shortcut avoidance. Agents that fake a passing test or skip a hard step are a known headache. If Carbon really does this less, long autonomous runs become more trustworthy.
- Prompt injection resistance. An agent that reads web pages, files, and repos for days will run into a lot of hostile input. Resistance to injected instructions is a basic requirement for that kind of work.
The halving of unsupported claims also matters here. In a multi-day task, one confident fabrication can derail everything that follows.
📊 Competitive Picture
The frontier is moving fast. Google shipped Argon earlier this month and is already replacing it. Carbon now leads both Anthropic’s and OpenAI’s flagship models on a widely watched index.
Keeping Argon’s introductory price is a deliberate move. Google is offering a better model at the same cost, which puts pressure on rivals’ pricing. One leaderboard doesn’t settle the race, though. Real-world performance on your own workloads is what counts, and most teams can’t test Carbon yet.
📋 Recommended Actions
- Security teams: Find out whether your vendors are in Google’s early-access group. If they are, ask how they’re using Carbon.
- Engineering leads: Plan for agents that can run for days. Decide now what review gates and permissions those long runs will need.
- Budget owners: Note the $2/$10 pricing. It’s a useful baseline when you compare frontier model costs.
- Everyone else: Watch for the Fairwind Program expansion. That’s when broader access starts.
🔭 Outlook
Carbon shows where frontier competition is heading: longer autonomous work, stronger security skills, and more careful decisions about who gets access first. Expect rivals to answer with their own security-focused releases and staged rollouts. Full details are in the original report from The Verge AI.