Google launches Gemini 4 Carbon, first for vetted cyber defenders
Google announced Gemini 4 Carbon on 28 October 2026, calling it its strongest model for cybersecurity and long-running tasks. Access starts with a small group of vetted cyber defenders through the Fairwind Program. Google has not set a date for developers, businesses or consumers. According to Google, the model ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and it scores 88.2% on SEC-bench Pro for finding software vulnerabilities. The introductory API price is $2 per million input tokens and $10 per million output tokens. For a business that automates work, the practical point is simple: the model is not yet open to most companies, so planning should start with the tools you already have.
In short
- Google announced Gemini 4 Carbon on 28 October 2026, with access first limited to vetted cyber defenders.
- Google reports a score of 59 on Artificial Analysis Intelligence Index v5.0, the top position it cites.
- Developers, businesses and consumers have no firm date for access.
- The introductory API price is $2 per million input tokens and $10 per million output tokens.
- Security benchmark scores were measured with guardrails disabled, so they do not describe the customer version.
What happened
On 28 October 2026, Google DeepMind published a post announcing Gemini 4 Carbon, which the company calls its next frontier model. According to Google's own announcement, the model shows gains in coding and agentic tasks, and it leads several security tests. Koray Kavukcuoglu, Google's chief AI architect, wrote the post. Carbon follows Gemini 4 Argon, which Google released about two weeks earlier, on 30 September 2026.
Google's headline figures come from its own benchmark table. The company reports a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index, placing it first. It also reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Google ran the security tests with guardrails disabled, so these scores describe the model's raw capability rather than the version customers will use.
Why it matters for businesses that automate work
Most companies that automate routine work use general-purpose models for document handling, data checks and reports. Carbon's launch matters for two reasons. First, Google is positioning the model around long, multi-step jobs, where the system keeps working on one problem for a long period. Second, Google is restricting the first release to people it has vetted. That signals that the company treats its strongest cyber capabilities as a controlled release, not a general product.
Google also said it is working with the U.S. government through a voluntary pre-release testing process. Other governments have not, as far as the announcement states, set a similar step. Businesses in any country should expect the rollout to follow that review.
What changes in practice
For most businesses, nothing changes this week. Carbon is not available to general customers, and Google has not published a date. The introductory price of $2 per million input tokens and $10 per million output tokens is listed for the model, but it does not apply to your account until you can use it. Pricing after the introductory period is higher, according to AI Weekly.
The Fairwind Program is the route for cyber teams. Google's Fairwind page describes a pairing of Gemini 3.8 Flash Cyber with its CodeMender tool, which finds, verifies and fixes vulnerabilities. Google says any Google Cloud customer can use CodeMender with publicly available models, which is an option you can review now, separate from Carbon.
| Question | Current position, October 2026 |
|---|---|
| Who can use Carbon | A subset of vetted cyber defenders, starting the week of announcement |
| Developers and businesses | No firm date; Google says access will come once guardrails are refined |
| Introductory API price | $2 per million input tokens, $10 per million output tokens |
| Independent index score | 59 on Artificial Analysis v5.0, per Google and independent reporting |
Two checks matter before any budget decision. Benchmark scores from one index version cannot be compared with scores from another, so the 59 figure should not be set against older Argon results. Also, the long-running capability is described in Google's own internal testing and has not been independently verified in the public record.
What to watch next
Three signals will show whether Carbon becomes useful to ordinary operations teams. The first is a public date for developer or enterprise access. The second is whether Google publishes the final, guardrailed version's scores, since the security figures were measured without guardrails. The third is whether the Fairwind rules widen beyond vetted organisations. Until then, the model is a story about the frontier of AI, not a tool to switch on in your workflow.
Automation is not a replacement for a security team. Models that find and patch flaws still need a person to approve changes to live systems, check the results against your own records and decide what gets deployed.
Sources
How AiStaffo would automate this
AiStaffo would start by listing which routine tasks in your business touch software, such as vulnerability alerts, vendor patch notices and access reviews. The system would connect to the systems that hold that information, such as your ticketing tool and asset list, and run a scheduled check that sorts new alerts by the software you actually use. A named person in your team would receive a summary and approve any change before it reaches live systems. You still decide what gets patched, who gets access and when a model is trusted with more work. Book a free automation audit.
Questions people ask
What is Gemini 4 Carbon?
Can my business use Gemini 4 Carbon now?
How much does Gemini 4 Carbon cost?
What is the Fairwind Program?
Does a higher benchmark score mean a better tool for my operations team?
Book a free automation audit
Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.
