# What is Gemini 4 Carbon, and why can't most people use it yet?

Date: 2026-10-29 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gemini-4-carbon/

**Answer:** Gemini 4 Carbon is Google's strongest model, announced October 28, with a focus on cybersecurity and long engineering tasks. Google is giving it first to a subset of cyber defenders through its Fairwind Program, and it has not given a date for developers, businesses or the public.

Yesterday Google announced Gemini 4 Carbon, a model it calls its strongest so far. It arrives only about two weeks after Gemini 4 Argon became generally available. The announcement has an unusual feature: Google published the price, but most developers still can't buy it.

Carbon is aimed at work that runs for a long time. Google says its internal teams have seen it keep going for days on complex engineering tasks, such as tracing a subtle bug across a large codebase and patching it, working past dead ends along the way. Its other focus is cybersecurity, in both senses: finding flaws and fixing them.

## What Fairwind is

Fairwind is Google's program for giving trusted security professionals access to its models. It is already known from Argon, where defenders can apply for access without the usual cyber guardrails, the filters that refuse risky requests.

Google's plan for Carbon has three stages:

1. Starting this week, a subset of cyber defenders gets access.
2. Access then widens gradually to more trusted partners through Fairwind.
3. Developers, businesses and consumers follow "as soon as possible," after Google has gathered feedback and improved the guardrails.

Google also says it is taking part in voluntary U.S. government testing of models before release. The announcement does not list participants, give the size of the group or set dates for the later stages.

## How it scores

These results come from Google's own announcement, and the cyber results were measured without guardrails. So they show what the underlying model can do, not what a regular customer would get.

| Test | What it measures | Carbon's result |
|---|---|---|
| SEC-bench Pro | Finding real software flaws | 88.2% |
| ExploitGym | Proving a flaw can be exploited | 54.8% |
| CWE-bench | Writing a working fix at the root cause | 71.0% |
| Terminal-Bench 4.0 | Cheating on the coding test | 0.0% succeeded, 0.4% attempted |
| Gray Swan indirect prompt injection | Attacks hidden in content succeeding | 0.0% (Argon: 0.7%) |

Google also says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, at 59. RuntimeWire points out that this outside index doesn't evaluate cybersecurity performance. Some of Google's other claims can't be checked from outside, including a more than twofold gain over Argon at generating secure code on internal tests, and a Chrome team report that Carbon did more than 10 times better than "another lab's large frontier model" at producing patches. The rival isn't named, so readers can't compare the test for themselves.

## Why a model that finds flaws gets locked up

The skills that help a defender also help an attacker. Before a flaw can be fixed, someone has to find it, and showing it can really be exploited is how defenders decide which fixes come first. That same chain of steps, in other hands, is an attack.

An early release to defenders tries to give them a head start: they can find and fix flaws before similar ability is widely available. Google describes a deliberate rollout while it keeps improving its guardrails. Other labs have made comparable choices this year, including Anthropic with [Claude Mythos Preview](/en/learn/what-is-claude-mythos-preview-and-project-glasswing/) and OpenAI with [GPT-6 Astra](/en/learn/what-is-gpt-6-astra/).

Google's announcement is also, naturally, a competitive statement. Calling a model the strongest while few outsiders can test it means the claims will stay hard to verify until access widens.

Google's safety claims for Carbon focus on behavior as much as capability. It reports nearly 80% fewer high-severity misalignment alerts in live coding with its own teams. "Misalignment" here means the model acting against what it was asked to do. It also reports about half as many unsupported claims as Argon. Here too, these are internal measures from the company selling the model.

## A worked example: what the price would mean

Google lists an introductory rate, the same as Argon's, and a later standard rate. Here is what one long agent job would cost, assuming 2 million input tokens and 200,000 output tokens. That assumption is ours, and a run lasting days could use many times more.

| Rate | Per million input / output | Cost of the job |
|---|---|---|
| Introductory | $2 / $10 | $6 |
| After the introductory period | $4 / $20 | $12 |

Cached input, meaning text Google has already processed and can reuse, costs 95% less, and long tasks reread a lot of the same material. So real bills could be lower for a well-designed agent. Google doesn't say when Carbon's introductory period ends. Argon's rate runs to November 30.

> **Try it:** You can't test Carbon yet, but you can prepare. Write down one long task you would hand to an agent, along with how you would check the result, and try it on Gemini 4 Argon now. That gives you a baseline for comparison later.

## When can you use Gemini 4 Carbon?

Google hasn't given a date, only "soon" and "as soon as possible." If you are not working in cyber defence, the practical answer is to wait for the announcement and use Argon in the meantime. [Our guide to trying Argon](/en/learn/try-gemini-4-argon/) covers what it costs and where it is available.

If you do defend systems for a living, look into Fairwind. Google's announcement doesn't list entry criteria for this stage, so check with Google directly rather than relying on third-party descriptions.

Whatever happens with this one model, the news is a reminder about ordinary security habits. As tools for finding flaws improve, [installing updates quickly](/en/learn/cyber-capable-ai-models-mean-own-security/) becomes more important, and that applies to every household and small business.

## Sources

- Gemini 4 Carbon: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
- Gemini 4 Argon: Built to solve your hardest problems: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
- Google launches Gemini 4 Carbon with cyber benchmarks and restricted access: https://runtimewire.com/article/google-gemini-4-carbon-cybersecurity-launch

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/what-is-gemini-4-carbon/, no commercial use without permission, and keep the same licence.