Gemini 4 Carbon: who can access it, and what will it cost?

Carbon has been announced. When can ordinary users access it?

Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect, announced Carbon on October 28, 2026: a plan to begin access for selected cyber defenders during that week, followed by a gradual expansion to trusted Fairwind partners. Broader developer, business and consumer access is forthcoming.

Separate three questions: what Google has announced, whether your product shows the model, and whether your account can actually use it. A launch story does not establish the latter two. Seeing a model name is no reason to retire an existing, working process prematurely.

Who is Fairwind for, and what does applying mean?

Fairwind prioritizes governments and national cyber authorities, critical infrastructure operators and core technology platforms. It checks organizations' security and ethical track records. Requests go through a Google Cloud account team, with no fixed approval turnaround. At this review, the technology section and FAQ still explicitly describe Argon; they do not establish Carbon access for every partner.

Prepare a defensive use case, authorized data and system scope, internal users and access controls, then confirm the approved model and conditions with Google. Matching a priority category, submitting an application, receiving approval and obtaining working access are different stages. Do not treat one as completion of all four or buy purported Fairwind access from a reseller.

How should you budget for the introductory prices?

Carbon's announced launch rates are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward. Cached input is priced 95% below the input rate. Google gives no Carbon introductory pricing end date.

Tokens measure content processed by a model; they are not a fixed word count or a number of requests. Track input, output and qualifying cached input separately, then check your channel's actual billing rules. A cached-input discount does not reduce the whole request, its output or a subscription fee by 95%.

For example, a task that repeatedly reads context and produces several responses cannot be costed from its first input alone. Set a task budget and stopping conditions, then measure the full consumption after receiving access. No Carbon calls or actual bills were generated for this article.

Do long-task capabilities justify an immediate migration?

Google positions Carbon for cyber defense and sustained, multi-step software engineering, citing internal use and evaluations. Those are vendor reports, rather than evidence of your task success rate, a guaranteed runtime or a security guarantee.

After receiving access, start with a small task whose scope is clear. Fix acceptance criteria, a budget, tool permissions and human review points. Record code output, test results and cost separately. After an interruption or uncertain result, check the original task before deciding to retry.

Keep a tested model available as a working fallback. Before migrating, confirm the official entry point, exact model identifier, request documentation and account limits. Do not guess API parameters from a headline or treat a different network exit as a way to obtain eligibility.

What is the most useful next step today?

Ordinary users can follow updates to the Carbon announcement and the release notes for their actual product before deciding whether to change plans. Upgrading for unconfirmed model access does not guarantee the result you want.

Organizations can confirm their cohort through their existing Cloud account team. Obtain the model, entry point, use conditions and billing details before scheduling a pilot. Leave unspecified countries, quotas and dates explicitly unconfirmed, and use evidence from a small trial to decide the next step.