Gemini 4 Neon shipped Nov 18 with a ski sim built from shaders, physics and spatial audio. What it changes for designers, the hype vs reality, and who decides.
Google released Gemini 4 Neon on November 18, 2026, and the demo designers are passing around is not a benchmark chart. It is a first-person ski simulator: WebGL2, real Mont Blanc terrain, 10,200 lines of custom shaders, 120 Hz physics and spatial Web Audio, written by the model in Google Antigravity with no pre-made assets. Gemini 4 Neon is the mid-tier frontier model of the Gemini 4 family, priced at $3 per million input tokens and $15 per million output, with a 1 million token context. My verdict after reading the official launch: Neon makes the expensive layers of an interface (shaders, physics, sound) cheap to try. It does not decide which of those layers your product should have, and that call is still the designer's job.
Below: what shipped, what it changes for people who design interaction and motion, the hype vs reality check, and where I would put it in the way I already build.
All facts here come from Google's official Gemini 4 Neon announcement. Google describes Neon as frontier-level intelligence with fast end-to-end task completion for software engineering, knowledge work and science. It sits next to Gemini 4 Hydrogen, the fast and cheaper model Google shipped a week earlier, and Google says Neon can act as an orchestrator that directs parallel Hydrogen subagents inside Antigravity.

The parts that matter for a designer:
Last week I wrote about the Gemini front end design skill that shipped with Hydrogen. Hydrogen was about a better-looking first page. Neon is about something heavier: long, tool-heavy builds that keep going without falling apart.
Read the ski demo again as a designer, not as an engineer. Shaders are the look. Physics is the feel. Spatial audio is the sound. Those are three layers that, until recently, needed three specialists or one very patient person. A model that can write all three in one session moves the cost of trying them close to zero.
That matters to me because the layers in that demo are the same ones in my own work. My Haribo pick and mix experiment uses three.js for rendering, Rapier for physics, a custom jelly shader for the bears, and sound synthesized in the browser. My stdo. keyboard landing page is built around Web Audio synthesis, where every key plays a sound. I built those with Figma, Claude and the Figma MCP. The hard part was never typing the shader. It was deciding how heavy a gummy bear should feel and how loud a key should be.
So here is what I think actually changes this month:
The demos are impressive, so it is worth being precise about what they prove.
Every showcase is Google's own, inside Google's own harness. The announcement itself shows how much the harness matters: on SWE-sweep, Neon scores 12.8% in the mini-SWE harness and 39.0% in Google Antigravity. Same model, three times the result. If you test Neon in a different tool, expect something different from the launch videos.
The benchmarks are reported by Google. Several are independent indexes, but the fabrication number comes from internal testing. Treat it as a direction, not a guarantee, and still check what the agent claims it did.
A ski simulator is not a product. It has no onboarding, no empty states, no error states, no accessibility settings, no account, no paywall. It is a wow moment. Products are mostly the boring screens around the wow moment, and that is where most of my design time at ViewStats went. When creators signed up but did not connect their channel, the fix was not a better shader. It was showing value first and asking for trust second, which took channel connection from 13% to 65%+.
More layers can make things worse. Physics, shaders and spatial audio on every screen is noise. The skill is choosing the one moment that deserves them.
Access costs money. Antigravity, where the showcase builds ran, needs a paid AI Pro or Ultra plan. The API is pay per token.
| Task | What Gemini 4 Neon can take on | What still needs a designer |
|---|---|---|
| Shaders and visual effects | Write long custom shader code from a description | Decide the look, the restraint, and where an effect earns its place |
| Physics and motion | Implement real-time physics at high frame rates | Tune weight, friction and timing until it feels right in the hand |
| Sound | Generate spatial Web Audio with no audio files | Choose which moments get sound and how quiet the rest stays |
| Product screens | Build layouts, states and components quickly | Define onboarding, trust moments, empty and error states from real users |
| Corrections mid-build | Adapt to changes without restarting (its #2 steerability claim) | Know what to change, which comes from looking at the build with taste |
| Verification | Report progress and claim tasks are done | Check it on a real device, with real content, against the design |
I wrote on X about how I stopped handing off designs: collect references first, design only two moments, talk to Claude like a developer friend, and change the design mid-build. A model ranked high on steerability is built for that last step. That is where I would test Neon first, not on a one-shot prompt.
Concretely, I would use it in three places:
1. The wow moment, after it is designed. I still design the hero interaction in Figma first, then hand the frame to the model as the brief. Neon's strength in shaders and physics fits the part of an experiment that a design file cannot express. The frame keeps the model honest about layout and type.
2. Orchestrated builds. Google pitches Neon as an orchestrator directing Hydrogen subagents. For a landing page with a 3D hero, a sound layer and a set of plain sections, splitting the work like that makes sense: one smart planner, several fast builders. I would keep the planner on a short leash and review each piece.
3. Sound passes. Most product teams skip sound because nobody owns it. A model that writes spatial audio cheaply means I can try a sound on a confirmation, hear it, and cut it if it is annoying. Trying is the hard part; cutting is easy.
Where I would not use it: deciding what to build. No model told me that creators wanted confidence more than charts. That came from interviews, Discord and onboarding data. If you want the longer version of this argument, I wrote about the design taste AI can't copy and why there is no magic prompt in vibe coding.
Gemini 4 Neon is a Google model released on November 18, 2026, positioned for complex software engineering, knowledge work and science. It has a 1 million token context window, four thinking levels, and costs $3 per million input tokens and $15 per million output tokens, according to Google.
It is good at the code side of rich interfaces: Google's demos include custom WebGL shaders, real-time physics and spatial audio. It does not replace the design decisions about structure, states and trust. Use it to build what you designed, not to decide what to design.
Hydrogen is the faster, cheaper model and shipped with Google's front end design skill, so it suits quick layout exploration. Neon is the stronger model for long, tool-heavy builds and mid-task corrections. Google suggests combining them, with Neon directing Hydrogen subagents.
Google lists the Gemini API in Google AI Studio, Gemini for business, the Gemini app for Pro and Ultra subscribers, and Google Antigravity for paid AI Pro and Ultra subscribers.
They replace a lot of typing. They do not replace knowing which interaction deserves physics, how a sound should feel, or which screen loses users. That judgment is what a design engineer is hired for.
If you design interfaces, pick one moment in your product that would benefit from weight, motion or sound. Design it in Figma first. Then build it with a model, change it twice mid-build, and test it on a real phone. You will learn more about what Neon can do from that one moment than from any benchmark.
If you want to see what this looks like when it ships, browse my AI design experiments, each one a working build. And if you are building an AI product that needs someone who designs the interaction, the motion and the sound, and ships the code, work with me as your product designer.
If you have a web or mobile app idea in mind or you need some advice about product design, feel free to get in touch ✉️