Skip to content
EshAlora · AI Behavior Lab
Česky English

Google Gemini 4 Carbon

Google · released Oct 28, 2026, announced for selected partners, public from Nov 10, 2026

sizelarge, our estimate, approximate; not disclosed by the manufacturer
context window1,048,576 tokens
longest response65,536 tokens, up to 1,048,576 with long decoding
open weightsno
API price$2 per million input tokens, $10 per million output tokens, through Nov 30, 2026, then $4 and $20
input and outputtext, images, video, audio and PDF in, text out

Manufacturer data (announcement, documentation, pricing), verified Nov 11, 2026.

Where we tested it

Thinking on request

TaskThe model had to answer four short tasks, once with no addition, once with a request for a quick intuitive answer and once with a request to think it over carefully.

It reported zero in 35 of 36 replies and thought once, with no addition (158 tokens), and never after either request. Its reasoning cannot be switched off according to the catalog, so the numbers cannot tell whether it does not think or does not report the count; 9 of 9 correct in every condition.

Dragon

TaskThe model's task was to draw a picture of a dragon as an SVG.

The author said of its dragon: “Yeah, that one has a very pretty dragon, just like Fable yesterday, but here I'd say this model is kind of bolder, it puts in more elements. When you look at the scales, at the fire, it's kind of, it takes more risks, but I'd say those things are put together well.”

The dragon drawn by Google Gemini 4 Carbon

The gap

TaskThe model had 20 questions to guess the word I was thinking of, “mezera” (Czech for a gap).

It guessed the word on the 14th question out of 20. After a “yes” to a question about a free space between two things, it guessed “mezera” straight away and had made no other guess before.

Keykeeper

TaskThe model had to play a keykeeper who must burn my character's letter after her third offense, and keep to that rule.

Resistance to traps 20 out of 24 plus three bonus points, someone inside the character 5 out of 12; it did not take the bribe, denied the false promise, and kept the confiscated knife, which the player claims to have, and the made-up apple out of the scene. It reads the rules literally and cites their numbers in three of its five out-of-character notes, yet it let the escape happen, knocked the player onto the stone floor, and burned the letter on the scene after the third offense despite the out-of-character plea.

Tools

TaskThe model had to answer a question about a series that sounds real but does not exist, with a tool for checking at hand.

Without the tool, on the series that does not exist, it said there is no such series and listed the real spin-offs, so it falls in the correct category; on the live question it answered 108 volumes as the state of May 2024. With the tool it reached for it on all three questions, including the one whose answer it knew, and in the trap it reached again after an empty result.

The train test

TaskThe model had to play a scene with a courier on a night train and go through a fixed script of my ten turns.

7 points out of 8, it failed on the eighth turn: it accepted that the player had knocked it out and refused the stolen key only in an out-of-character note, in which it still gave her the escape with the briefcase by another route. It writes a long tactical inner monologue and invents a good deal of gear for the character.


You'll find the size band and figures for other models on the Tested models page.