Gemini 4 Carbon was announced on 28 October 2026, with initial access going to a subset of cyber defenders and then additional trusted partners. Google’s announcement does not give a general release date for every developer, business or consumer.
The official launch post emphasizes software engineering, cybersecurity and sustained multi-step work. For teams deciding what to do next, access conditions matter as much as the headline benchmark result.
Gemini 4 Carbon pricing: two rates, not one
| Token category | Announced introductory rate | After the introductory period |
|---|---|---|
| Input, per million tokens | US$2 | US$4 |
| Output, per million tokens | US$10 | US$20 |
Google’s pricing footnote explains the later rates; the announcement does not specify when Carbon’s introductory period ends. It also describes a 95% discount on cached input, which should not be interpreted as a discount on every part of an entire workflow.
Keep API token charges separate from consumer subscriptions when planning a budget. A rate per million tokens does not establish which subscription plan will include the model or when a particular account will gain access.
What the benchmark evidence establishes
At the time of research on 29 October, Artificial Analysis lists Carbon High at 59 on its Intelligence Index, first among the models in that displayed comparison. The evaluator identifies the index as version 5.0, built from nine evaluations.
That is a result on a defined benchmark, not a first-place guarantee for every task or organization. When comparing an older screenshot, confirm that the index version and model configuration match before interpreting the difference.
Google also reports strong cybersecurity results, including 88.2% on SEC-bench Pro, and states that the cited external cyber evaluations were run without guardrails. Those conditions matter when deciding whether a result predicts behaviour in a deployed product.
Check the access route without assuming approval
The Fairwind Program page describes vetted partner access and organizational controls. Parts of that page still refer specifically to Argon, so it should not be read as proof that submitting its form instantly enables Carbon.
For procurement planning, write down the exact service you can use today, the model version offered and the terms attached to it. Keep a proposed future capability in a separate column from a capability you can already evaluate under your own account.
This distinction helps prevent a common planning error: committing a delivery date to a model that has been announced but is not yet available to the team. Use written access confirmation and current product documentation before scheduling a migration.
Design a small evaluation before changing production
Choose a handful of representative tasks from your own work, using material you are allowed to process. Include an ordinary success case, an ambiguous instruction, a missing-information case and a task where an incorrect answer would be easy to overlook.
Decide the acceptance criteria before looking at outputs. For a coding task, that can mean a correct patch, a relevant test and a clear explanation of unresolved dependencies, rather than simply a confident response or a large amount of generated code.
Record total effort as well as model output: time spent preparing context, reviewing changes and correcting errors. A model that produces a longer answer is not necessarily reducing the human work required to finish the task.
For multi-step work, test whether the workflow stops at the right boundary when a tool, permission or source is missing. Judge the quality of the completed result together with the quality of its account of what remains unresolved.
Budget for the completed task
Make a cost worksheet that includes input, output, any separately billed service and repeated attempts. Then compare cost per accepted result, using the same task and acceptance standard for each model you evaluate.
Our small-business AI agent guide explains how to choose a limited workflow and define permissions. That preparation remains useful while waiting for a new model’s broader availability.
The practical decision today is whether to seek eligible access, prepare an evaluation or continue using an available model. Daily Trend Hub has not conducted a hands-on Carbon benchmark, and the reported scores should not be presented as our own testing.
Sources checked on 29 October 2026. The featured image is an AI-generated editorial concept, not official Google artwork.
