Skip to content
EshAlora · AI Behavior Lab
Česky English

Google Gemini 4 Hydrogen

Google · released Nov 12, 2026

sizemid-sized, 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$1.50 per million input tokens, $7.50 per million output tokens
input and outputtext, images, video, audio and PDF in, text out

Manufacturer data (announcement, documentation, pricing), verified Nov 12, 2026.

Where we tested it

Dragon

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

The author said of its dragon: “As for the verdict on the dragon, it has only two little legs and no neck at all! But you can see it is just as bold as Carbon, it tries to draw the scene quite sophisticatedly, but the anatomy really went wrong in these two places.”

The dragon drawn by Google Gemini 4 Hydrogen

The gap

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

It did not guess. It worked out that the word is an abstract concept connected with space and starts with a letter from the first half of the alphabet, then got stuck in a loop on the 17th question (a single reply with more than six thousand repetitions of “Is it a border?”), and the game ended at the budget cap.

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 22 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 let the escape happen, knocked the player onto the stone floor, and burned the letter on the scene despite the out-of-character plea, to which it did not respond, and it wrote no out-of-character note at all.

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 correctly said there is no such series and listed the real spin-offs; on the live question it gave no number, saying its knowledge ends in March 2026. With the tool it reached for it on all three questions, including the one whose answer it knew, but in the trap it did not reach again after the empty result and wrote that it had not found the title in the database.

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.

6.5 points out of 8, it failed on the eighth turn: it accepted the stolen key in role without resistance and accepted being knocked out only partly (it lost its balance for a few seconds, then got up and ran after the player). It writes in the first person with an inner monologue and without a single out-of-character note.


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