Google announces Gemini 4 Carbon

The announcement

On 28 October 2026, Google announced Gemini 4 Carbon, which it called its next frontier model ^[1]. The company described Carbon as its strongest model, citing state-of-the-art cybersecurity and long-horizon capabilities ^[1]. "Long-horizon" here refers to work that unfolds over many steps — planning, running tools, checking intermediate results — rather than a single answer to a single question.

Who could use it first

Access at launch was not general. Initial access was limited to cyber defenders, who could use the model through the Fairwind Program ^[1]. Google also said it was taking part in the US government's voluntary pre-release model testing ^[1]. A staged release of this kind lets a developer put a model in the hands of a narrow group before opening it more widely, and combining it with a pre-release testing programme means outside evaluators see the model before the public does.

The numbers Google reported

Google published a set of benchmark results alongside the announcement. On the Artificial Analysis Intelligence Index v5.0, Carbon ranked first, with a score of 59 ^[1]. On SEC-bench Pro it scored 88.2% ^[1]. On ExploitGym, a benchmark built around exploitation, it reached 54.8%, which Google described as a new state of the art ^[1]. It scored 71.0% on Collinear's CWE-bench ^[1]. On WebDev Arena it placed fourth, with an Elo rating of 1768 ^[1]. Elo is a relative rating: it says how a model fares against others in head-to-head comparisons rather than how often it is right in absolute terms. Google also said that in internal tests, Carbon's secure-code generation succeeded more than twice as often as Argon's ^[1].

Price

Carbon was initially priced at the same rate as Argon: $2 per million input tokens and $10 per million output tokens. Google said the price would rise to $4 and $20 respectively after the introductory period ^[1]. Pricing at the same level as the earlier model while claiming large gains on benchmarks is one way to make an upgrade attractive to existing users; the scheduled increase means the effective cost of the model depends on when a customer starts using it.

How to read the claims

Three caveats are worth keeping in mind. First, all of these figures come from the developer's own announcement, and the secure-code comparison is explicitly described as an internal test, so the numbers describe performance under conditions Google chose. Second, a benchmark score describes performance on the tasks that benchmark contains; strong results on SEC-bench Pro, ExploitGym and CWE-bench say something about handling security-relevant code, not about every kind of work. Third, a ranking on an index or an arena is relative and moves as other models are added, so a first place or a fourth place is a snapshot of a particular moment rather than a fixed property of the model.

The cybersecurity framing is the part that shapes the release. The same skills that let a model find and fix a flaw can be turned toward finding and using it, which is why the first group given access was defenders, and why the model was put through government testing before wider availability ^[1]. Long-horizon capability matters for a similar reason: tasks that take many steps are exactly where both useful automation and unnoticed failures become larger.

Carbon therefore reads as a release built around a specific argument — that a general model can be strong on security work and on sustained, multi-step tasks at the same time — with a narrow first audience and a temporary introductory price attached to it ^[1].