Notes on film, new tools and the decisions behind the work.

Unconfirmed

“Krypton” takes twice as long in a claimed Gemini coding test, creator says

Chetaslua says an alleged Krypton checkpoint took roughly twice as long and wrote twice as much code as Carbon and Argon on one prompt. The model’s identity and origin remain unverified.

A 12-second video shows three versions of a 3D goose riding a bicycle, each labelled with a different alleged model: Gemini Krypton, Carbon or Argon. Chetaslua (@chetaslua) says the three were given the same prompt; in the post, the creator calls the Krypton version a first checkpoint, “krypton-a”.

X / DEMO

Chetaslua’s three-way bicycle goose comparison

Loading post from X…

Credit / Chetaslua / @chetasluaOpen original 

The model identities are creator-supplied labels and have not been independently verified.

The creator says the Krypton-labelled run took roughly twice as long and produced about twice as much code. That describes one reported attempt, not a repeatable benchmark: more code does not by itself show better results, and the post does not disclose where the supposed model ran or establish its origin.

Google’s public model catalogue lists Gemini 4 Carbon and Argon but does not list Krypton as of 11 November. Absence from that catalogue is not proof that no internal model exists. Google has not confirmed the name in the sources we checked, and the video does not authenticate the model behind any of its labels.

Sources

  1. Three labelled goose-on-a-bicycle demos attributed to Krypton, Carbon and ArgonChetaslua / X · 9 Nov 2026
  2. Gemini API models catalogueGoogle AI for Developers

Earliest source found: . Updated: .