Gemini 4 Argon tested against GPT-6 Astra

Robbert van Empel published a practical comparison of Gemini 4 Argon and GPT-6 Astra on October 13, 2026. His video examines Google's new model through two creative coding briefs: redesigning Wikipedia's toilet paper orientation article and building an interactive LEGO globe with a South Park theme.

The test follows Google's broader rollout on October 13. Argon is available through the paid Gemini API, to Google AI Pro and Ultra subscribers in the Gemini app, and to Ultra subscribers in Antigravity. Google lists introductory API prices of $2 per million input tokens and $10 per million output tokens until November 30. Those rates describe token usage; the final cost depends on the work a model performs.

The intelligence chart shown in the video gives Argon and GPT-6 Astra the same score of 53 points. That makes the practical comparison interesting: an aggregate score measures a range of abilities, while a finished website or game also depends on design decisions, animation, controls and implementation.

In The Future's website evaluation, Argon scores 73 out of 100 and places fourteenth. Its newspaper-inspired design combines serif typography, patent illustrations, a quiz and a household calculator. Several interactions work, but the central 3D toilet paper roll falls outside the camera's view. Mobile layout and accessibility defects further reduce the execution score.

The LEGO game evaluation is stronger, placing Argon fifth with 78 out of 100. The world includes character conversations, a main quest, a side mission and movable objects. Humor and a playable storyline give it personality. The assessment also identifies building collisions that allow players partly into walls, skipped dialogue and progress that disappears after reloading.

For makers, these results suggest evaluating the entire output of a coding task. Visual appeal, working controls and the effort needed to repair defects all influence usefulness. A shared prompt helps compare submissions, while these scores remain assessments of two specific projects. The embedded video shows Van Empel exploring the results, and the linked benchmark pages let readers inspect the designs and play the game themselves.