Google stages Gemini 4 Carbon access for cyber defenders

An announcement with a narrow first audience

Google announced Gemini 4 Carbon on October 28, saying access would begin that week with selected cyber defenders and expand through its Fairwind Program. Wider developer, business and consumer availability has no firm date.

That is a staged release. It should not be read as an invitation for any developer to build against a generally available model today.

What Fairwind access means

The program’s public description prioritizes governments, critical infrastructure operators and core technology platforms. Applicants undergo due-diligence checks; approval is not automatic.

Its governance rules limit work to defensive and research purposes, require access controls and prohibit partners from reselling or redistributing model access. Those are published program conditions, not an independent audit of how every participant operates.

The overview still names Gemini 4 Argon in its technology and access sections. Carbon’s rollout is established by the new announcement; the older program page should not be treated as a complete Carbon-specific service contract.

Read the price footnote

Google lists launch prices of $2 per million input tokens and $10 per million output tokens. Its footnote says these rise to $4 and $20 after the introductory period, without specifying when that period ends.

Tokens meter what the model reads and generates. A price per token is not a complete project budget: the amount of work a task consumes still matters.

Why the cyber scores need their test conditions

Google reports strong cyber benchmark results, explicitly measured without guardrails. These are capability claims under particular test conditions, not measurements of the full restricted service.

One named benchmark, SEC-bench Pro, tests agents on vulnerability discovery and proof-of-concept generation. Its public overview describes a containerized test setup, time limits and automated checking. Its overall table aggregates 344 instances across V8, Firefox and Linux.

At this report’s October 29 check, the public page did not list Carbon. It therefore does not independently reproduce Google’s Carbon result. That absence is not proof the company’s evaluation is wrong; it limits what can be verified from that table.

Comparisons need the same benchmark version, task set, tools, budget and safeguards. A score from one configuration cannot establish the reliability or safety of a different deployment. Model Current has not tested Carbon or run these benchmarks.