Google expanded Gemini 4 Carbon access on November 10, opening paid routes beyond its initial cybersecurity rollout.

The big change

  • Who can use it: Paid developers can use Carbon through AI Studio or Android Studio; businesses have Gemini for business. Ultra subscribers are gaining access in the Gemini app, Search's AI Mode and Antigravity, Google's coding environment.
  • Why it matters: Google says it designed Carbon to work through difficult engineering problems over multiple steps. Following a bug across several files requires more than producing plausible code: proposed changes still need to be checked.
  • What to watch: Artificial Analysis measured different token use at Carbon's High and Medium reasoning settings. The cost of a completed task therefore depends on how much work the model performs, not just its advertised token price.

The earlier rollout

Google's October 28 introduction offered Carbon first to selected cybersecurity defenders through its Fairwind program, with trusted partners to follow. Google said it was gathering feedback and refining safeguards before broader distribution.

Researchers can apply for six months' access, subject to eligibility, capacity and selection.

An application programming interface, or API, connects a model to other software. A developer sends requests from an application and receives output that the application can use.

Google described internal work tracing and patching subtle bugs across large codebases. To understand that workflow, imagine a developer asking where a fault originates, then checking a suggested repair against the surrounding code. That is an illustration of the stated use, not a test by BIG CHANGE.

Google's model card warns that Carbon can invent incorrect information, become slow or time out. Its specialist cyber capabilities remain restricted to verified defenders.

What the price means

Tokens are the units used to meter model input and output. A request and its response can contain different quantities, so both sides contribute to the bill.

For public-model API use, Google's introductory rates are $2 per million input tokens and $10 per million output tokens through November 30. The rates then double.

One million input tokens plus 100,000 output tokens would therefore cost $3 initially and $6 later, before other applicable charges. This illustration is not a customer quote.

What independent measurements show

Artificial Analysis compares Carbon's High and Medium settings across its Intelligence Index, a combination of nine evaluations. It reports weighted averages of 123,000 and 103,000 output tokens per task respectively. The higher setting used more output tokens in that test mix; it does not follow that every individual job will behave the same way.

The evaluator also runs Google's Multimodal Knowledge Agent benchmark independently. Its 200 tasks ask agents to turn visually dense files into deliverables such as reports and spreadsheets. A common software environment gives the models tools, and an AI judge checks the submitted work against weighted criteria. This tests a defined workflow, not whether every customer's final document is correct.

BIG CHANGE has not independently tested Carbon. Test representative jobs whose results you can check, recording token use and the time needed to review and correct the output.