Gemini 4 Argon Built a 3D Periodic Table in Antigravity: How to Recreate It

Google DeepMind has a new favorite demo. On October 20, 2026, the lab posted a 1:28 video of Gemini 4 Argon building a 3D periodic table in Antigravity that animates orbiting electrons, atomic data and light spectra, and syncs to Tom Lehrer's song "The Elements." The post passed 319,000 views within about a day.

It is a charming demo and a useful one for builders. This post covers what Google actually claimed, what the model and tool cost, how the build likely works, and a prompt plan you can run yourself. One limit up front: Google published the video but not the prompt, the code or the number of iterations, so anything below about how it was built is our reconstruction, labeled as such.

TL;DR: questions builders ask first

What did Google actually say?

The post from Google DeepMind is two sentences: Argon "built a 3D periodic table synced to Tom Lehrer's song, 'The Elements.'" It was "created in Antigravity," and "animates orbiting electrons, atomic data, and light spectra." Third-party sites covering it say the visuals match both the beat and the lyrics, but those reports rely on the same video.

Context matters. Argon was announced on September 30 for Fairwind testers, as we covered in our launch post, and Google's launch announcement says it reached developers and enterprises through the API on October 13, 2026. The announcement highlights spatial reasoning, mathematical precision and visual coding, specifically for 3D visualizations and interactive apps. A periodic table demo is a direct showcase of that claim.

What does Argon cost and who can use it?

From Google's launch post:

A build like this is a good fit for the intro pricing window, because iterating on visuals burns output tokens. For the full benchmark and price picture, see our Argon comparison with Opus 5.5, Grok 4.7 and GPT-6 Astra.

How would you build it? Our reconstruction

Treat this as an informed guess, not Google's recipe. A song-synced 3D element table breaks into five parts.

Math tutor triangle and ruler on a desk, illustrating an interactive learning app such as an AI-built periodic table

1. Data. All 118 elements with atomic number, symbol, mass, shell configuration and emission lines. Ask the model to load a vetted dataset instead of recalling values, then spot-check a few entries. Models can slip on obscure numbers.
2. 3D scene. A browser build in Three.js or WebGL: a grid of element tiles, a camera that flies between them, and a detail view with orbiting electrons per shell.
3. Spectra. Each element's emission lines drawn as colored bars, from wavelengths in the dataset.
4. Timeline. A cue sheet mapping song timestamps to element IDs, so the camera lands on each element as it is named.
5. Audio. An audio element with a play button, driving the timeline from the playback clock.

Antigravity's agent loop, slash commands such as /goal and custom skills, per Google's launch notes, suit this kind of multi-file build. Our guides to Antigravity Interactive Generative UI Artifacts and Antigravity Plan Mode show how to structure it.

A prompt plan you can run

Split the work so the model checks each piece before the next.

Goal: a single-page 3D periodic table in the browser (Three.js, no build step).
Step 1: write elements.json for all 118 elements from a source I will name.
        Fields: number, symbol, name, mass, shells, emission_lines_nm.
Step 2: render a 3D grid of tiles. Click a tile for a detail view with
        orbiting electrons by shell and an emission-line bar.
Step 3: add a guided tour mode with play, pause and next, driven by a
        timeline array of {time_s, element_number}.
Step 4: wire the timeline to an audio file I supply. Pause must stop both.
After each step, run it, screenshot it, and list what is wrong before fixing.

Two choices make a difference. Ask for screenshots and self-critique after each step, since visual bugs hide from code review. And keep the timeline as data in its own file, so you can retime it without regenerating the scene.

How do you keep the build on track?

Visual projects fail in predictable ways, and each has a cheap guard. These are our suggestions from general practice with agentic builds, not steps Google described.

Pin the data first. Have the agent write the element file in step one, then read ten random rows yourself against a trusted reference. Wrong atomic masses or shell counts are easy to miss once they are baked into animations. Lock the file afterwards so later steps cannot quietly rewrite it.

Keep the scene and the timeline separate. The scene renders any element on request. The timeline is a plain list of times and element numbers. When the song cue drifts by a second, you edit a number, not a prompt. This is also the part most likely to need a human ear.

Ask for a debug overlay. A small on-screen readout of the current time, element and camera target makes sync bugs obvious in a screenshot. Without it, you are guessing why the camera landed on the wrong tile.

Test on a laptop, not just your workstation. A 3D scene with 118 tiles, particle electrons and a spectrum panel can stutter on integrated graphics. Ask the agent for a low-detail mode and a frame-rate cap, and turn the electron count down when many tiles are in view.

Handle accessibility. A guided tour that only works with sound excludes some learners. Add captions or on-screen text for each element, keyboard controls for play, pause and next, and a reduced-motion setting that stops the orbits. These are small requests to a model and make the result usable in a classroom.

What are people asking about the demo?

Is the tour interactive or just a video? Google's description covers animation and sync. One reply asked for pause-and-compare behavior on the spectra, which suggests viewers could not tell. Build that in yourself if you want it: a click on a tile should halt the tour and open a side-by-side spectrum view.

Does it need Antigravity, or will any agentic IDE work? Nothing in a periodic table requires a specific tool. Antigravity is where Google made the demo, and its agent loop and skills suit multi-file work. The same prompt plan works with other coding agents; expect differences in how they handle screenshots and long sessions.

How big is the cost? At the introductory $2 and $10 per million tokens, a project like this is likely to cost a few dollars in model fees if you iterate sensibly, though that is our estimate and depends on how much output you regenerate. The bigger cost is your review time. After November 30 the rates double, which is a reason to do exploratory builds sooner.

Can a teacher use it as is? Not without checking. Verify the science, add captions, and confirm the audio is licensed for classroom or public use. Treat any AI-generated teaching tool as a draft that a subject expert signs off on.

Why does a song-synced demo matter at all? It tests several skills at once: structured data, 3D geometry, timing and polish. A model that handles all four in one session is more convincing than a single-purpose benchmark score, even though the demo cannot tell you how often it works.

What should you be skeptical about?

• It is one clip. We do not know how many attempts it took. A reply under the demo from a developer said they got a zen garden from Argon "with one prompt" in under five minutes, but that is also anecdote. A Carlos E. Perez post nearby said it is "still a struggle to get Argon to conjure up inspiring visuals," so reports are mixed.
• Interactivity is unknown. One reply asked whether you can pause on an element and compare its spectrum with a neighbor. The video as described is a guided tour; whether it supports free exploration is not stated.
• Accuracy is unknown. Atomic data and spectra were not independently checked by us. For anything you ship to students, verify against a reference.
• Music rights. Syncing a copyrighted recording or lyrics to a public app raises licensing questions. We do not know how Google handled it in the demo. Use licensed or original audio, and do not reproduce lyrics in your build.
• Tool access. Antigravity access depends on your plan; check what your tier includes before planning a project around it.

What this means for what you build

The useful lesson is not the song. It is that a frontier model plus an agentic IDE can turn a one-sentence idea into a polished teaching tool in a sitting. One X reply said they envy "all these kids growing up with ... a personal tutor," which captures why education is a natural target. If you teach or train, an interactive explainer for your own subject is now a small project, not a quarter-long one.

Pair it with a human check. Treat the model as a fast front-end engineer and your subject expertise as the reviewer. Our list of top Gemini 4 Argon use cases shows other builds worth studying. If you let an agent run code on your machine to do this, review what it executes; AgentBeam, from the explainx.ai team, stops AI agents before they take dangerous actions.

Claimed versus verified

Related reading on explainx.ai

• Gemini 4 Argon launch, benchmarks and pricing
• Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
• Top 10 Gemini 4 Argon use cases
• Antigravity Interactive Generative UI Artifacts
• Antigravity Plan Mode slash command
• Antigravity CLI: sandbox, plugins and subagents
• Gemini 4 Argon for robotics: ERQA 87
• Official: Gemini 4 Argon announcement

Pricing, availability and access tiers are accurate as of October 21, 2026 and may change.