Gemini 4 Neon codes a WebGL ski game as Antigravity agents build and play mini-golf
By GameSlash
Google's Nov 18–19 Neon demonstrations show a browser ski simulator, a multi-agent mini-golf build and AI browser playtesting. Source videos, workflow and caveats.
Google introduced Gemini 4 Neon on November 18, 2026, and posted a first-person Chamonix ski simulator example on November 19. The game-development angle is concrete: in Google Antigravity, the model wrote a browser-based WebGL2 ski project using real Mont Blanc terrain, without pre-made 3D meshes or sound assets. Google says the demo combines around 10,200 lines of custom shaders, 120 Hz ski physics, dynamic snow tracks, triggered slab avalanches and spatial audio across six routes. These are Google's reported project figures, not independent GameSlash measurements. Google launch announcement · Google's original video post.

A second demo: a game that lets the coding agent join as Player 2
Google's Antigravity release post describes using `/swarm` to scaffold a multi-level, two-player game in the built-in browser while an agent takes the second player's role. Frames of the official NeonPutt '84 demonstration show a task graph with a coordinator and workers for physics, shaders, UI, courses and verification. A later frame shows the neon mini-golf course running at `localhost:5173` and a browser-playing agent waiting to take its shot after the human player.

The second-player screenshot identifies the browser-playing executor as Gemini 4 Hydrogen. That distinction matters: the developer's example uses Neon to coordinate or construct the project, and another model can handle a narrower browser-use task. Google has added an Orchestrator option that explicitly pairs Neon for planning and review with Hydrogen for coding, builds and tests, returning control to Neon when a task needs replanning. Antigravity model-orchestration details.

What to copy in an AI-assisted game workflow
Rather than ask an agent for a whole large game, give it a small testable scene: one course or ski slope, explicit controls, one fail/restart condition and a measurable check for collision behavior. Delegate art/shaders, gameplay physics, interface and verification separately only when the project is large enough to benefit. Then have the agent open the actual browser build and test interactions while a human verifies the feel, frame pacing and unexpected outcomes. Google's example illustrates an approach; it is not evidence every generated build will pass a gameplay QA review.
Access, costs and what is not available
Gemini 4 Neon is rolling out in the paid Gemini API and to eligible Google AI Pro/Ultra users in Antigravity. The API list price in Google's announcement is $3 per million input tokens and $15 per million output tokens (cached input $0.15 per million). Those API prices do not automatically describe Antigravity subscription quotas. Google's demo videos are viewable, but the announcement did not publish playable builds or source repositories for the named ski and mini-golf examples. So this article does not claim that readers can download or play them. See the original launch note and the Antigravity game demo description.