AI & Machine Learning

Google announces Gemini 4 Carbon. Cyber defenders get it first, everyone else waits

2 min read

Charcoal-style illustration of a faceted carbon lattice rising from smudged charcoal dust, with electric-blue pathways branching from its nodes, for Google's launch of the Gemini 4 Carbon AI model.

Google has announced Gemini 4 Carbon, a frontier model it pitches at cybersecurity and long-running engineering work. Access starts with a small group of cyber defenders, and Google gives no date for wider release.

In brief

This story is developing and will be updated.

On 28 October 2026 Google announced Gemini 4 Carbon, which it calls its "strongest model", built for cybersecurity defence and long-horizon engineering tasks. It follows Gemini 4 Argon, which Google says launched two weeks earlier.

Who gets it, and when

Google says that from this week Carbon goes to "a subset of cyber defenders", with access then phased to more trusted partners through its Fairwind Program. Developers, businesses and consumers get it later, "as soon as possible", after more guardrail work; Google gives no date or regions. Google also says it is taking part in the US government's voluntary pre-release model testing process.

Pricing

Google says Carbon will launch at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period, $4 and $20 apply. Google does not say when that period ends for Carbon. ByteNib's Argon pricing brief covers how Argon's window works.

How it differs from Argon

All the following are Google's claims, not independent results:

  • Security benchmarks. Google reports state-of-the-art scores of 88.2% on SEC-bench Pro (finding vulnerabilities) and 54.8% on ExploitGym (exploitation), plus 71.0% on CWE-bench (patching), all evaluated without guardrails.
  • Secure code. On internal benchmarks, more than twice Argon's success rate at generating secure code.
  • Long tasks. Internal teams saw Carbon run for days on multi-step engineering work.
  • Behaviour. In Google's internal testing, roughly half as many unsupported claims as Argon, and 0.0% attack success on Gray Swan's indirect prompt-injection benchmark, against Argon's 0.7%.

What it means for IT leaders

Most organisations cannot use Carbon yet, so plan around Argon for now. Security teams should ask whether they qualify for early access. A model that Google says can find and exploit vulnerabilities also changes the threat model: assume attackers will have similar tools.

Sources and scope

Capabilities, access, pricing and benchmark figures are from Google's announcement (28 October 2026), checked the same day. Implications are ByteNib editorial analysis.

Drafted with AI assistance and reviewed by the editor before publication.

Continue exploring: AI & Machine Learning analysis, the related implementation tutorial, and the structured learning path.