Google introduced Gemini 4 Carbon on October 28, targeting cybersecurity and coding jobs that run for days. Its first announced users are selected cyber defenders. If you're wondering whether another impressive model score means you should change your everyday AI setup, access comes before the comparison.
Carbon's launch puts two audiences in the same announcement. Security teams want earlier help finding and fixing flaws. Everyone else wants to know when the model will reach the products they use. Google has described the first rollout much more precisely than the second.
Who gets the first turn
The Carbon announcement starts with a subset of cyber defenders, expanding through DeepMind's Fairwind Program. The broader release is described as coming soon, without a public date in that announcement.
Fairwind prioritizes organizations such as governments and critical infrastructure operators. Its application route is through a Google Cloud account team. That is a different route from choosing a model in an ordinary chat app.
We've followed this staged approach in the Gemini 4 Argon story, which includes its later Cloud release. The lesson for a buyer is fairly mundane: establish which service you can actually obtain before comparing the headline capabilities.
What the cybersecurity scores measure
Google reports 88.2% on SEC-bench Pro for finding vulnerabilities and 71.0% on CWE-bench for patching. It specifies that the external cybersecurity evaluations were "evaluated without guardrails". These are reported benchmark results, not a measurement of every task allowed through the eventual customer service.
That qualification changes a practical comparison. A model can demonstrate that it knows how to complete a security task while the service restricts who can request it. A team buying defensive tooling needs to know both what the model can do and which of those actions its access agreement permits.
The patching result is especially worth following. A useful finding has to become a change someone can maintain. We'd want to see how a patch behaves with the project's tests, how it handles an awkward dependency and whether its explanation helps the next engineer. A benchmark score alone doesn't answer those questions.
The announced price has a second chapter
Carbon's announced introductory API rates are $2 per million input tokens and $10 per million output tokens. After the introductory period, the stated rates are $4 and $20. Google's Carbon post doesn't give an end date for that introductory period, so don't borrow one from another model's offer.
For work that runs over several days, the amount of context reread and the number of attempts also shape the bill. A lower rate is useful; a job that repeatedly circles back can still cost more than expected. The eventual service needs to make that spending visible while work is underway.
Carbon gives defenders a concrete model to evaluate and other customers a release to watch. Before assigning it a budget, ask for the access terms and the price that will apply after the welcome offer ends.
Cover: AI-generated editorial illustration by TMRW using Google branding. The code-review pages and security key are illustrative, not Carbon output or a product screenshot.
