Google put out its newest frontier model, Gemini 4 Carbon, on Wednesday, and the company is leading with one claim above all others: this is its most capable model yet at finding and fixing security flaws.
Carbon now sits at the top of the Artificial Analysis Intelligence Index (v5.0), a widely watched benchmark, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, according to Google. It also holds the number one spot on two narrower tests: SEC-bench Pro, which measures how well a model can find real bugs in software, and ExploitGym, which checks whether a model can turn a discovered vulnerability into a working exploit.
That's the pitch. Carbon isn't a general release, not yet. Google says it's currently available only to a select group of cybersecurity defense companies, with access expanding over time to other partners through its Fairwind Program. There's no public timeline for when everyday developers or consumer-facing products get it.
What's actually new
Carbon follows Gemini 4 Argon, which Google shipped earlier this month, and the company is framing the gap between the two as proof of how fast it's moving. Google says Carbon brings a "significant leap" in cyberdefense, getting better at spotting vulnerabilities and writing code that doesn't introduce new ones.
The model is also built to grind through longer jobs. Google says its internal teams have watched Carbon "run for days to persist through dead ends to tackle complex, multi-step engineering tasks," rather than giving up or taking shortcuts partway through. Google also says it's more resistant to prompt injection, the trick where malicious text hidden in a webpage or document tries to hijack an AI agent's instructions, and better at working alongside a human user instead of around them.
On honesty, Google claims Carbon cuts what it calls "unsupported claims" by half compared with Argon. That's Google's own framing of a chronic problem with large language models: confidently stating things that aren't true.
Pricing stays put. Carbon launches at the same introductory rate Argon shipped at, $2 per million input tokens and $10 per million output tokens, the unit AI companies use to meter how much text a model reads and generates.
The bigger story here is positioning, not just capability. Cybersecurity has become the arena where Google, OpenAI and Anthropic are all racing to prove their models can defend networks faster than attackers can break them, and a model that tops bug-hunting and exploit-generation benchmarks is a loud opening argument. Whether that translates into real-world defense, outside of a curated group of partner companies, is the thing nobody outside Google can check yet.
For now, the rest of us wait on the Fairwind Program to widen.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.