Gemini 4 Carbon.

One of the smartest of the 58, but one of the most expensive and slowest per task.

Google · Closed weights · Released 10 November 2026
Data from 13 November 2026.

  • Score2nd of 58

    60

    Artificial Analysis's Intelligence Index. The median is 27.

  • Cost43rd of 51

    $5.48

    What each task of that index costs. The median is $1.03.

  • Time38th of 50

    28.4 min

    How long each task of that index takes. The median is 14.6 min.

  • Context5th of 55

    1.05 million

    Tokens it reads at once, as 8 others do.

In short

It scores 60 on Artificial Analysis's Intelligence Index, 2nd of the 58 models it scores. The median is 27.

Each task of that index costs it $5.48 and takes it 28.4 minutes on average. The median of the 58 is $1.03 and 14.6 minutes.

Google sells it at $2 per million input tokens and $10 per million output tokens, and from 1 December, $4 and $20, and it reads up to 1.05 million tokens at once.

Its weights are not published.

Its data

Maker
Google
Weights
Closed
Effort measured
High
Price per million tokens
$2 input and $10 outputFrom 1 December, $4 and $20
Context
1.05 million tokens
Released
10 Nov 2026

The score, the cost and the time per task and its tests from Artificial Analysis, read on 13 November 2026 at its highest effort. The price, the context and the release date from Google's page. The table of the 58.

Its tests

Coding

  • FrontierCode 2.0 Changes to real open-⁠source projects. 7th of 58.40.6%
  • Terminal-⁠Bench 4.0 Complex work in the terminal, on its own. 2nd of 58.65.2%

Reasoning

  • HLE Diamond Expert-⁠level questions. 2nd of 58.69.1%
  • CritPt v2, hard subset Physics research problems. 4th of 58.69.0%

Questions.

  • How much does Gemini 4 Carbon cost?

    Google sells it at $2 per million input tokens and $10 per million output tokens, and from 1 December, $4 and $20. On the tasks of Artificial Analysis's index it costs $5.48 each.

  • How much text does it read at once?

    Up to 1.05 million tokens, according to Google.

  • Can it run on your own servers?

    No. Its weights are not published, so it is used through Google or whoever offers it.

  • Which one for your company's software?

    It depends on the job. We cover it in “Which AI Model to Use to Build Your Company's Software”.

Let's talk.

Tell us what you need and one of the partners who write the code gets back to you.