Carbon was announced on 28 October 2026 and is not a general developer, business, or consumer release. Scores and the secure-code comparison are Google’s reported results. Introductory API prices are scheduled to rise, and Google has not stated when.
1. What Google announced
Google describes Carbon as its strongest Gemini model for cybersecurity and long-horizon work, building on Gemini 4 Argon from about two weeks earlier. The company says researchers and models are pushing capability and safeguards together, and that Carbon is in the U.S. government’s voluntary pre-release testing process.
The rollout is deliberate. Google said it would give access this week to a subset of cyber defenders, then add trusted partners through the Fairwind Program. Broader access for developers, businesses, and consumers comes after Google gathers feedback and adjusts guardrails. Until Carbon appears in a product you already pay for, you do not have a model to switch to.
2. The claims, and what they do not prove
On small screens, scroll the table sideways to see every column.
| Claim | How to read it |
|---|---|
| Artificial Analysis Intelligence Index v5.0 score of 59, ranked first | A third-party index as reported by Google. It is not a score on your codebase or your ERP. |
| SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% | Cybersecurity benchmarks for finding issues, showing they can be used, and patching. They are not permission to test systems you do not own. |
| More than twice Argon’s success at generating secure code, on an internal benchmark | An internal comparison with the previous Gemini model. Ask for the task set before you treat it as a purchasing fact. |
| Long-running engineering tasks | Google says internal teams have seen the model persist through dead ends. Duration is not the same as a correct change. |
Google also says Carbon cuts unsupported claims relative to Argon and is more resistant to prompt injection. Those are product claims to retest on your own prompts. A model that is better at finding vulnerabilities is a sharper tool. Use it on code and systems you are allowed to assess, and keep a person responsible for any change that ships.
3. Price after you actually get access
Google says Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. After the introductory period, the price is $4 per million input tokens and $20 per million output tokens. The announcement does not give the date that introductory period ends.
Do not put the introductory rate into a year-long cost model. When access arrives, record the price shown in your account and the date you checked it. Cached-input discounts only help if your workload actually repeats a large prompt.
4. What to do while access is closed
- Leave production on the model you can call today. A benchmark lead you cannot invoke does not speed up a release.
- Write the eval now. Save one coding task and one business task with a known good result. When Carbon appears, run those same tasks before you change a default.
- Separate cyber defense from general office work. The first audience is cyber defenders in Fairwind. An invoice workflow or a helpdesk draft may wait for the general release.
- Keep testing inside your boundary. Vulnerability scores are not a how-to for breaking into a customer, partner, or public system.
If your team already uses Gemini for drafting or code review, keep that setup and note Carbon as a later comparison. Switching tools because a press post ranked a model first is how a project loses a week.
5. Frequently asked questions
Can my business call Gemini 4 Carbon today?
Not as a general release. Google said the first access goes to a subset of cyber defenders through the Fairwind Program, with developers, businesses, and consumers to follow after further guardrail work.
Is the $2 per million input price permanent?
No. Google calls that rate introductory, matching Gemini 4 Argon, and says the later price is $4 per million input tokens and $20 per million output tokens. The announcement does not say when the introductory period ends.
Should we replace our current coding model because Carbon ranks first?
Only after you can run it on a task you can check. A leaderboard score, including Google’s reported first-place Intelligence Index result, does not show that Carbon is available in your account or better at your repository.