Google announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest model for cybersecurity work. It is not publicly available: access starts with a select group of cyber defenders and widens in stages. The benchmark scores come from Google and have not been independently verified.
Contents
show
What Google says
According to SecurityWeek (October 29):
- Access: opened first to selected cyber defenders, then to approved partners through Google’s Fairwind Program, which serves governments, critical infrastructure operators and major technology platforms. Google says more than 650 organisations participate.
- Lineage: Carbon follows Gemini 4 Argon, launched two weeks earlier.
- Reported benchmark results: 88.2% on SEC-bench Pro (finding flaws in real software), 54.8% on ExploitGym (exploitation) and 71.0% on Collinear’s CWE-bench (patching).
- Comparison: Google says Carbon beat OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in its own comparison, with results obtained without guardrails.
- Pricing: introductory rates of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period.
- Broader launch: Google says it will gather feedback and refine guardrails before offering it to developers, businesses and consumers, with no date.
How to read the claims
- Vendor-run tests. The scores were reported by Google. Without independent replication, treat them as marketing-grade evidence.
- “No guardrails” cuts both ways. Evaluating a model without safeguards shows raw capability, but it is not what vetted users will get.
- Access is the story. Several leading labs now hold their strongest security-capable models back from general release. Anthropic does the same with Mythos 5, which is limited to verified organisations; see our explainer on Claude Fable 5 and Mythos 5 and the report on government approval of Mythos.
What it means for most people
Nothing changes for ordinary Gemini users today. The practical effect is for security teams: if a vendor in your supply chain is a Fairwind partner, it may begin patching faster, but that is an expectation rather than a guarantee. We have not seen independent tests, so claims such as Carbon being more than twice as successful as Argon at writing secure code should be read as Google’s internal result.
