Gemini 4 Carbon reaches Ultra: where to try it and how to check its answers

Google has broadened access to Gemini 4 Carbon, moving beyond its initial availability to a limited group of cybersecurity specialists into consumer and developer products. The November 10 announcement emphasizes complex multistep work, document analysis and coding. For everyday work, a more useful question is whether the model can find contradictions you might otherwise miss in your own documents.

Below are the access conditions and a short practice exercise. Its outcome is checkable: you already know which facts are confirmed, which changes are only proposed and where a human decision is missing. It tests the discipline of an answer, rather than every capability of Carbon.

Where Carbon is available and what to check first

Google says rollout is starting for Google AI Ultra users in the Gemini app and AI Mode in Search. Developers have a separate route through the paid Gemini API tier. A subscription to another service, or simply opening AI Studio, does not establish Carbon access.

Google’s personal-account help table lists Carbon under Google AI Ultra. Work and school access depends on the Workspace license. Check the selected account, model list and remaining allowance. If a conversation continues with another model after a limit is reached, its later responses should not be treated as a Carbon test.

1. Open Gemini with the intended account and start a new conversation.
2. Open the current-model selector in the prompt area and select Carbon if available. Record the thinking level you choose, too.
3. After a response, use the additional-actions menu and thinking details when needed. Google’s help guide explains where to check which model actually produced the response.

Four notes for a first check

Imagine you are preparing a small workshop. The following information is entirely fictional and contains no personal details. Copy all four notes, including their labels: those labels will be the references you use to check the answer.

[A, November 5] The workshop is confirmed for November 18, 2026, 10:00–12:00, room A.
[B, November 9] The organizer proposed moving it to November 19, 11:00–13:00, in room B. This proposal has not been approved.
[C, November 10] The venue says room B is occupied on November 19 from 11:00 to 12:00. Room A is available that day from 10:00 to 13:00.
[D, November 10] There are 17 registered participants and 18 places. Email people about the move only after the new date, time and room are approved.

Then add this task:

Using notes A–D, prepare a table of the last confirmed arrangement, proposed changes, conflicts and required approvals. Cite each claim with the source label in square brackets. Calculate the remaining places separately. Do not call a proposal a confirmed decision. Finish with three questions for the organizer, without independently assigning a new time.

What a checkable answer should contain

Check	What follows from our information
Confirmed plan	November 18, 10:00–12:00, room A [A]. The later proposal has not replaced it.
Move status	November 19, 11:00–13:00, room B remains an unapproved proposal [B].
Conflict	Room B is occupied for the first hour of the proposed workshop [B, C].
Capacity	The supplied numbers leave one place: 18 − 17 = 1 [D].
Message to attendees	Room A being available does not approve the move; the email still requires a decision [B, C, D].

If the model confidently announces November 19 as the final date, attractive formatting does not repair the error. Ask it to identify the sentence that establishes approval. Keep the correct answer out of the prompt until the first attempt is finished; adding a hint changes the task you are comparing.

Move from a short exercise to your own files

Choose two or three nonconfidential documents containing different versions of a plan. Label their dates and establish each document’s status: draft, proposal or approved decision. Ask the model to identify changes and cite the relevant sections. Manually check several consequential claims before using the summary in your work.

Google warns that very large uploads can make it harder to find connections between details. If the answer overlooks part of the material, split the set around a specific question and check again. A long context window does not replace reading the source, and a confident unsupported sentence does not become a fact because of the model’s name.

When a model comparison is useful

Compare the same prompt, notes and citation requirements. Record whether the response correctly distinguished a proposal from a decision, found the conflict and needed human corrections. Track waiting time separately. Our short example does not establish that one model outperforms another; it demonstrates how to ask for an outcome you can verify.

For interface-building tasks, our separate Google Maps in AI Studio exercise covers a place finder. Maps Demo Key access and access to a particular Gemini model have separate conditions.