Gemini 4 Argon launched broadly on 13 October 2026, and it currently ranks first on the Vals Index. So I did what I always do with a new frontier model: I built five apps and scored each one out of 5.
Two scored 5 out of 5, one scored 3 and two scored 2. In this review I show you every app, give you the prompts to rebuild them, and compare my results with Google's claims, independent leaderboards and real posts from people using the model.
Key Takeaways
Gemini 4 Argon at a Glance
Here are the facts you need before the tests. Everything in this table comes from Google's own pages.
| Item | What Google says |
|---|---|
| Announced | 30 September 2026, first to vetted defenders and partners through the Fairwind program |
| Broad release | 13 October 2026, when Google said it was available globally |
| API model ID | gemini-4-argon (model page) |
| Thinking levels | Low, medium (default) and high (thinking docs) |
| Context window | 1,048,576 tokens, shared by input, thinking and output |
| Output length | 65,536 tokens by default, with a long decode mode for longer runs |
| Inputs | Text, image, video, audio and PDF |
| Outputs | Text only (no image or audio generation) |
| Knowledge cutoff | March 2026 (model card) |
| API price | $2 input and $10 output per 1M tokens through 30 November, then $4 and $20 (pricing page) |
How I Tested Gemini 4 Argon
I wrote one prompt for each of five apps, built them with Argon and scored each finished app out of 5. The five apps cover the kinds of work people actually ask a frontier model to do: a 3D game, a 3D simulation, a multi page front end, a second 3D game with a stricter UI and an interactive tool with a timeline.
The scores are my own call on the finished app, as I show it in the video. I did not run repeat builds, so treat each score as a snapshot and not an average. I also did not test coding in a real repository, API behavior or long agent runs, and I say where other people's tests cover that.
Also checkout Gemini Prompt Generator here.
The Scoreboard
| # | App | What it is | Score | My verdict |
|---|---|---|---|---|
| 1 | Slipstream | 3D racing game | 5/5 | Neat graphics, intuitive controls, very fun |
| 2 | Event Horizon | 3D black hole simulator | 5/5 | Detailed, looks like a real black hole, UI is intuitive |
| 3 | Perfume house site | Six page front end | 2/5 | Spacing, card layout and scroll effects missed the brief |
| 4 | Harbour | Cozy 3D town builder | 2/5 | Nice animations, but controls and tablet layout break |
| 5 | Skyfire | Fireworks show designer | 3/5 | Works well, but not responsive |
| Total | 17/25 | Strong at 3D, weak at layout |
Demo 1: Slipstream, the 3D Racing Game (5/5)
I started with a racing game because it stresses everything at once: 3D rendering, game logic, menus and controls. The video says the model needed about 15 to 20 minutes to produce this build.
What Argon built
What I saw
The graphics look neat for a 3D game built this fast, and the controls feel intuitive. I am pretty bad at racing games, so I cannot judge handling or balance, but the game is fun to play and everything I clicked worked.
"Everything is great."
I gave it 5 out of 5. The result is similar to what I get from Fable and one other model, so Argon holds its own on this kind of build.
What other people report
Other people are building games with Argon too. One post describes turning photos of a home office desk into a playable racing game with Blender MCP and Three.js. Another compares Argon with Muse Wonder on the same game prompt, and the same account later claimed a cost of $29.71 for Argon against $3.45 for Muse Wonder. These are user claims, not audited numbers.
Prompt to rebuild it
This is my reconstruction from what the video shows. Your own prompt may differ.
Build "Slipstream" — a high-speed anti-gravity racing game in the spirit of classic futuristic
racers. Hovercraft scream around twisting neon tracks that loop, bank, corkscrew, and run
upside down. ONE self-contained slipstream.html. All geometry, textures, and audio procedural.
Modern three.js via the importmap above (three@0.170.0, SRGBColorSpace, ACESFilmicToneMapping,
no deprecated APIs). Addons allowed: EffectComposer, RenderPass, UnrealBloomPass, OutputPass,
SMAAPass.
═══════════ THE HOVER PHYSICS (get this right before anything else) ═══════════
Fixed timestep (240Hz) decoupled from render, interpolated for display.
- The craft floats on FOUR hover thrusters that raycast toward the track surface. Each applies
a spring-damper force to hold a target hover height. This is what gives the craft its
floaty, weighty feel — tilts on corners, bobs over seams, dips on landings.
- GRAVITY IS RELATIVE TO THE TRACK, not the world. Gravity points along the negative track
normal under the craft, so it sticks to loops, wall-rides, and inverted sections. When the
craft leaves the surface (a jump gap), switch to world gravity until it re-acquires.
Getting this transition right is the single hardest part of the build.
- The craft continuously re-orients so its "up" slerps toward the track normal.
- Thrust, drag, and a lateral grip that lets the craft slide wide in fast corners
- AIRBRAKES (left and right, Q/E): one-sided drag that pivots the craft hard into a corner.
Mastering airbrakes is the skill ceiling of the genre — make them feel essential.
- SLIPSTREAM DRAFTING: sitting in the wake cone behind an opponent reduces drag and fills a
boost meter. Visualize the wake as a faint distortion ribbon behind every craft.
- Boost pads on the track; a boost meter spent with Shift
- ENERGY / SHIELD: wall scrapes and collisions drain it; a pit lane each lap recharges it as
you fly through (slower than the racing line — a real strategic trade-off). Zero energy =
the craft explodes and respawns with a time penalty.
- Collisions with walls deflect along the wall rather than stopping dead; craft-to-craft
contact bumps both and costs energy.
═══════════ THE TRACKS ═══════════
Build every track from a 3D spline with a per-point ROLL (banking) and width, extruding the
road surface along parallel-transported frames so it can twist through a full 360° without
flipping. From the same spline derive: the AI racing line, checkpoints, the minimap path,
boost-pad positions, and the pit lane.
4 tracks, each visually distinct:
1. NEON SPRAWL — night city canyon between skyscrapers, emissive signage, a tunnel
2. DUST LINE — desert canyon at dusk, a long jump gap, a banked hairpin
3. HALO — an orbital ring station in space, a full loop and an inverted straight, Earth below
4. ICEFALL — glacial descent, a corkscrew, translucent ice walls with subsurface glow
Track surface: glowing edge lines, chevrons on corners, a scrolling energy texture on boost
pads, and trackside props instanced for performance.
═══════════ THE CRAFT ═══════════
6 teams, each a distinct procedurally built craft (fuselage, nacelles, fins, glowing engine
ports) with different stats: top speed, acceleration, handling, shield. Team liveries with
emissive accents. Engine trails as ribbon geometry that streak behind at speed.
═══════════ AI OPPONENTS (7 rivals) ═══════════
Follow the racing line with per-driver skill, brake (airbrake) based on upcoming curvature,
draft each other and the player, overtake by offsetting laterally, take the pit lane when
energy is low, and rubber-band mildly so races stay close. Difficulty: Novice / Pro / Elite.
═══════════ RACE STRUCTURE ═══════════
3-lap races with countdown, position tracking (lap + spline progress), lap and sector times
with deltas, wrong-way detection, respawn on falling off a jump. Results screen with an
animated standings reveal. Time trial mode with a GHOST of your best lap. A 4-race championship
with points. Speed classes (Venom / Flash / Rapier) that scale top speed — the top class
should feel genuinely frightening.
═══════════ CAMERAS (C to cycle) ═══════════
1. CHASE — spring-damped, rolls with the craft, swings wide in corners, pulls back at speed
2. COCKPIT — inside the canopy with a minimal holographic HUD projected on the glass
3. NOSE — low and fast, the most terrifying view
Plus an auto-cinematic REPLAY after each race with trackside camera cuts.
═══════════ THE FEEL (this genre lives or dies on sense of speed) ═══════════
- FOV widens with speed (eased), radial motion blur and speed lines scaling with velocity
- Bloom on engines, edge lines, and boost pads — tuned so it glows without washing out
- Chromatic aberration pulse on boost; screen shake on wall hits
- Sparks on wall scrapes; a heat-shimmer distortion behind engines
- AUDIO (Web Audio, synthesized): an engine whine that pitches with speed, a doppler-shifted
whoosh as rivals pass, airbrake hiss, boost roar, wall scrape, and a driving procedural
electronic soundtrack (kick, bass, arpeggio) whose intensity rises on the final lap.
═══════════ HUD & MENUS ═══════════
HUD: speed, energy bar, boost meter, lap, position, lap times with deltas, a minimap, and a
slipstream indicator when drafting. Menus (all animated): main menu with a rotating craft,
team select with stat bars, track select with preview flyovers, speed class, settings
(graphics quality Low→Ultra toggling post-FX and particle counts, controls, gamepad, volumes),
pause, results, championship standings. Gamepad support via the Gamepad API with analog
steering and trigger airbrakes. Persist bests and championship progress in localStorage.
═══════════ VERIFY ═══════════
□ The craft stays glued to the track through the full loop and the inverted straight
□ Leaving the surface on a jump switches to world gravity, and landing re-acquires cleanly
□ The track twists 360° without the extruded surface flipping or pinching
□ Airbrakes make tight corners possible that are impossible without them
□ Drafting measurably fills boost and AI uses it too
□ AI completes laps, pits when low, and overtakes without clipping through walls
□ Physics is identical at 30fps and 144fps (fixed timestep)
□ 60fps at High with 8 craft, trails, and full post-processing
Output the complete single HTML file, no truncation. Track-relative gravity, the surface
re-acquisition after jumps, and the twist-safe track extrusion are the critical systems.Demo 2: Event Horizon, the Black Hole Simulator (5/5)
My second demo asks for an interactive, physically grounded black hole simulator. You fly around the black hole and watch space time bend light in real time.
What Argon built
What I saw
The app looks pretty detailed, and the graphics look like a black hole as we imagine it or read about it. The controls respond when you raise or lower a setting, and the guided tour works. The UI is intuitive and fits the space theme.
"It does look like a space and a black hole."
I gave it 5 out of 5. I judged the physics by eye, so I cannot vouch for the equations.
What other people report
Google's own latest model page uses a Three.js black hole as a sample prompt, so this is a task Google expects people to try, though a sample prompt says nothing about output quality. On Hacker News, Simon Willison reported a Blender Python script for a Faberge egg that rendered well. That result points the same way as my test: Argon handles 3D scenes with confidence.
Prompt to rebuild it
Build "Event Horizon" — an interactive, physically grounded black hole simulator. Fly around
a black hole, watch spacetime bend light in real time, and drop probes toward the horizon.
ONE self-contained event-horizon.html. Modern three.js via the importmap above — but the core
is a full-screen fragment shader that ray-traces light through curved spacetime.
═══════════ THE PHYSICS (ray-traced in the fragment shader) ═══════════
Model a non-rotating (Schwarzschild) black hole in units where G = c = M = 1, so the event
horizon is at r = 2 (the Schwarzschild radius).
- For every pixel, trace a light ray BACKWARD from the camera, integrating its path through
curved spacetime. Use the standard real-time formulation that reproduces Schwarzschild null
geodesics: an acceleration term a = −1.5 · h² · r̂ / r⁵, where h is the ray's conserved
specific angular momentum |r × v|. Integrate with a stable step (RK4, or leapfrog with an
adaptive step that shrinks near the hole).
- If the ray crosses r < 2, the pixel is black (it fell in).
- If it escapes past a far radius, sample a procedural starfield/galaxy skybox in the ray's
final direction — so background stars visibly smear into arcs and an Einstein ring.
- If it crosses the ACCRETION DISK plane between the inner and outer radius, shade the disk.
THE ACCRETION DISK:
- A thin disk in the equatorial plane, from the ISCO (innermost stable circular orbit, r = 6)
out to r ≈ 20
- Temperature falling with radius (roughly T ∝ r^(−3/4)) mapped to blackbody colour — white-
blue hot inside, orange-red outside
- Animated turbulence: procedural noise advected around the disk at Keplerian speeds, so the
inner edge visibly rotates faster than the outer edge
- RELATIVISTIC DOPPLER BEAMING: the side of the disk rotating toward the camera is brighter and
bluer, the receding side dimmer and redder. This asymmetry is the signature of a real render.
- GRAVITATIONAL REDSHIFT: light from deep in the well is dimmed and reddened
What must EMERGE from the ray tracing rather than being painted on:
- The black SHADOW, with an apparent radius ≈ √27 ≈ 5.2 (larger than the horizon itself)
- The PHOTON RING, a thin bright ring at the shadow's edge
- The far side of the disk lensed UP over the top of the hole and DOWN under it — the
iconic image. If this doesn't appear, the geodesic integration is wrong.
PERFORMANCE: render the lensing pass at a configurable internal resolution (0.5×–1×) and
upscale; cap step counts; offer a quality preset. It must stay interactive on a laptop GPU.
═══════════ INTERACTION ═══════════
- FREE-FLY CAMERA: orbit, zoom, and fly. As you approach, the lensing distorts the whole sky.
- Orbit presets: a stable circular orbit at r = 10, the photon sphere (r = 3, where light
itself orbits — show it as a thin ring overlay), and a plunging trajectory
- Toggle the disk on/off to see pure lensing of the starfield
- Disk controls: inner/outer radius, temperature, turbulence, inclination relative to camera
- An "edge-on / face-on / 'cinematic' tilted" view preset set
═══════════ DROP A PROBE (the fun part) ═══════════
Launch a glowing probe from your position on a trajectory you aim. Integrate its MASSIVE-
particle geodesic (timelike, different from the light rays) so it orbits, precesses, or
plunges. Show:
- Its orbit trail, including periapsis precession for eccentric orbits (the rosette pattern)
- Its clock vs your clock: GRAVITATIONAL TIME DILATION factor √(1 − 2/r), displayed as two
ticking clocks side by side that visibly diverge as it falls
- As it nears the horizon it redshifts, dims, and appears to freeze and fade — it never visibly
crosses, which is exactly what a distant observer would see
- Probes inside the ISCO must spiral in; probes with enough angular momentum must orbit
═══════════ THE INSTRUMENT PANEL ═══════════
A clean scientific HUD: your distance from the centre in Schwarzschild radii, local time
dilation factor, escape velocity at your radius (as a fraction of c), orbital velocity, and
markers for the horizon (r=2), photon sphere (r=3), and ISCO (r=6). A short tooltip glossary
explaining each term in one sentence — this should teach as well as dazzle.
A guided "tour" mode that flies through 6 waypoints with a caption at each (the shadow, the
photon ring, Doppler beaming, lensed disk, time dilation, the ISCO).
═══════════ VISUAL POLISH ═══════════
Bloom on the disk and photon ring, ACES tone mapping, a rich procedural galaxy background
(Milky-Way-like band plus thousands of stars) so lensing has something beautiful to distort,
a subtle film grain. Screenshot button (canvas → PNG).
═══════════ VERIFY ═══════════
□ The far side of the disk appears lensed above AND below the shadow
□ The shadow's apparent radius is ≈ 5.2 (√27), not 2 — check against a measured overlay
□ A photon ring is visible at the shadow edge
□ One side of the disk is brighter (Doppler beaming) and the brighter side flips when you
orbit to the other side
□ Background stars form arcs and an Einstein ring when aligned behind the hole
□ A probe inside r = 6 spirals in; one with sufficient angular momentum orbits stably
□ The time-dilation clocks diverge correctly with radius
□ Interactive frame rate on a laptop at the default quality preset
Output the complete single HTML file, no truncation. The geodesic integration in the shader
is the whole build — if the lensed disk doesn't wrap over and under the shadow, nothing else
matters.Demo 3: The Perfume House Website (2/5)
This is the demo where Argon lost me. I asked for a complete website for an independent perfume house, made of six interlinked pages, and I defined each page in detail in the prompt.
What worked
I loved the interactive 3D bottle on the first screen. You can play with it, and the liquid inside looks smooth.
What went wrong
"This is bad, to be honest, from a company like Google."
The basic principles of UI design and user experience were not followed, so it gets 2 out of 5.
What other people report
My result lines up with other evidence, though not every post agrees.
| Source | What it says | Link |
|---|---|---|
| Arena WebDev | Argon ranks ninth at 1678 (preliminary, 1,772 votes), just behind GPT 6 Sol (max) at 1686 in eighth (read 13 Oct) | Leaderboard |
| Bloomberg, relayed by implicator.ai | Anonymous Google employees reportedly said Argon struggles with some coding tasks, including front end design, and that real world coding trails the benchmarks. Implicator summarizes the Bloomberg story second hand, and Google denies the underperformance claim | Report |
| @hqmank on X | A harder pelican on a bicycle front end test came out right, with correct geometry and small interactive touches | Post |
| @prathamdby on X | Mixed feelings on a personal front end test: some sections really good, some plain bad, and Opus 5.5 still the clear winner | Post |
| Claims "visual coding" gains in general terms, with no claim about responsive layouts or spacing on the pages I checked | Launch post |
Prompt to rebuild it
Build "Accord" — the website for an independent perfume house. A complete MULTI-PAGE site:
6 interlinked pages sharing one design system, one nav, and smooth page transitions. Sensual,
editorial, quiet luxury — closer to a fashion magazine than a shop.
STACK: three.js via the importmap above (three@0.170.0, modern API only), GSAP 3.12 +
ScrollTrigger, Lenis smooth scroll, Lucide icons, Google Fonts. All imagery procedural.
FILES:
shared.css · shared.js
index.html · collection.html · notes.html · discover.html · atelier.html · boutique.html
js/home.js · js/collection.js · js/notes.js · js/discover.js · js/atelier.js · js/boutique.js
DESIGN DIRECTION — "a scent made visible":
Light-first, with a dark "evening" theme. Palette: warm parchment, a deep aubergine ink, and
soft blush and amber accents that read as liquid colour. Fonts: a high-contrast editorial
italic serif for display, a light grotesk for body, and a small-caps mono for ingredient
names and concentrations. Generous whitespace, asymmetric editorial layouts, oversized type.
Slow, liquid easings. Nothing snaps.
THE 3D HERO (real Three.js WebGL — the centerpiece):
A heavy glass perfume FLACON with a faceted body and a weighted brass cap.
- The glass: MeshPhysicalMaterial with transmission, thickness, IOR ≈ 1.5, slight dispersion
tint at the edges, and an environment map so it refracts and reflects convincingly.
- THE LIQUID INSIDE: a separate mesh whose surface is a plane clipped to the bottle's interior,
tinted amber with its own transmission. It SLOSHES: the liquid surface tilts with a damped
spring driven by mouse movement and scroll velocity — fling the cursor and the perfume
swirls, then settles. This physical detail is what makes the hero feel alive.
- Fake caustics: a soft animated light pattern projected onto the surface beneath the bottle.
- Slow rotation; mouse parallax; dispose on page transition; gradient fallback without WebGL.
CUSTOM CURSOR (fresh): a soft mist cursor — a small blurred dot that leaves a faint drifting
trail of particles which diffuse and fade like an atomized spray. On hover over interactive
elements, a brief "spritz" burst. Hidden on touch devices.
PAGE 1 — index.html (11 sections):
1. Hero: the 3D flacon + house name + one line of positioning + two quiet CTAs
2. Press marquee (fashion and fragrance publications)
3. The house's three signature fragrances — tall editorial cards with hover tilt
4. THE PINNED SCROLL INTERLUDE (400vh) — "eight hours on the skin": a fragrance unfolds over
time. As the user scrolls, a clock advances from 0 to 8 hours. TOP NOTES (bergamot, pink
pepper) rise as bright particles and evaporate first; HEART NOTES (rose, iris) bloom in the
middle; BASE NOTES (amber, sandalwood, musk) settle low and linger to the end. Each note
names itself with a small-caps label as it appears. One continuous sequence, not slides.
5. The fragrance wheel teaser — an interactive olfactory families wheel → discover.html
6. Raw materials band: three ingredients, each with a procedurally illustrated botanical
7. Numbers (animated): raw materials in the palette, years of the house, months of maceration
8. A long-form editorial quote from the perfumer, set large
9. Atelier teaser → atelier.html
10. FAQ accordion (concentration, longevity, sampling, refills)
11. Final CTA + rich footer
PAGE 2 — collection.html: the fragrances. A grid where each fragrance has its own colour world;
filter by olfactory family, concentration (EDT / EDP / Extrait), and season; a detail view
with the full note pyramid, longevity and sillage meters, a "wear it when" mood line, sizes,
and a sample-set CTA.
PAGE 3 — notes.html: the ingredient library. A searchable, filterable index of raw materials
(naturals and molecules), each with a procedural illustration, origin, extraction method,
olfactory description, and which fragrances use it. A world map of where ingredients are
sourced, with hover cards.
PAGE 4 — discover.html: find your scent. An interactive OLFACTORY WHEEL (SVG — floral, oriental,
woody, fresh, and their sub-families) the user explores by hovering and clicking, plus a
6-question quiz (moods, places, textures, time of day) that scores answers against families
and recommends 2-3 fragrances with an explanation of why. Animated transitions between
questions; the result reveals with the matching flacon colour.
PAGE 5 — atelier.html: how it's made. The perfumer's story; a process timeline whose SVG line
draws on scroll (sourcing → extraction → composition → maceration → bottling); the organ (the
perfumer's desk of raw materials) as an illustrated interactive shelf; sustainability; footer.
PAGE 6 — boutique.html: visit and order. Boutique locations with hours, a refill programme,
a validated sample-request form (choose 3 fragrances, address, notes), gift wrapping, shipping
and returns, a closing full-bleed shot of the flacon. Footer.
SHARED SYSTEMS (shared.js): Lenis smooth scroll + a hairline progress bar; the mist cursor;
theme toggle persisted in localStorage; [data-reveal] entrance system with slow liquid easing;
subtle magnetic buttons; tilt cards with a soft sheen; a PAGE TRANSITION VEIL that dissolves
like mist rather than wiping.
REQUIREMENTS: 6 distinct background patterns (paper grain, soft radial bloom, fine botanical
line pattern, watercolour wash, dotted mist, silk gradient); active nav link; frosted nav after
scroll; mobile overlay menu; fully responsive; reduced-motion respected (the liquid stills and
the eight-hour sequence becomes static); accessible; 60fps; pixel ratio capped at 2. Note
pyramids and ingredient lists must be consistent across every page.
DELIVERY: shared.css and shared.js complete FIRST, then each page with its JS. No truncation.
End with a validation checklist.Demo 4: Harbour, the Cozy Town Builder (2/5)
Harbour is a cozy, goal free 3D town builder. It is a simple game where you build a harbor town block by block.
What worked
The animation looks good. The ships and the small details around the harbor add life to the scene, and the game is simple to play.
What went wrong
I do not like the controls. When I switch to the tablet device size, the layout is cut off or overflows to the left, so the app has no real responsiveness. For a game where you place blocks in a small area, broken controls hurt the whole experience.
I gave it 2 out of 5.
Prompt to rebuild it
Build "Harbour" — a cozy, goal-free town builder. The player clicks on water to raise colourful
buildings, and the town ASSEMBLES ITSELF: walls, roofs, arches, stairs, balconies, and gardens
appear automatically depending on what's next to what. ONE self-contained harbour.html.
Modern three.js via the importmap above. All geometry procedural.
═══════════ THE IRREGULAR GRID (the foundation — do this first) ═══════════
The charm comes from an ORGANIC grid, not a square one. Generate it:
1. Start from a hexagonal region subdivided into triangles
2. Randomly merge adjacent triangle pairs into quads (leaving some triangles)
3. Subdivide every triangle and quad into smaller QUADS so the whole grid is quads
4. RELAX the vertices iteratively (move each vertex toward the average of its neighbours, with
a force pulling each quad toward squareness) until it settles into a soft, organic mesh
The result: irregular four-sided cells that feel hand-drawn. Seedable, so a seed reproduces a grid.
═══════════ THE BUILDING SYSTEM ═══════════
Each grid cell can stack up to ~8 floors. Occupancy is stored per cell per height.
- DUAL-GRID MODULE SELECTION: for each corner point of the grid at each height, look at which
of the surrounding cells (above and below) are filled. That configuration selects a MODULE —
a piece of wall, a corner, a roof edge, an arch, a pillar, a floor edge.
- DEFORM each module to fit its irregular quad (bilinear mapping of the module's unit-square
geometry onto the quad's four corners) so every building conforms to the organic grid.
- Automatic architecture from context, at minimum:
* Peaked or sloped ROOFS on top of stacks, with chimneys occasionally
* ARCHES when a block hangs over an empty cell
* STAIRS climbing beside tall buildings
* Flat-roof GARDENS with tiny plants when a roof is wide enough
* BALCONIES and window boxes on some walls
* DOORS at ground level facing the water or a path
* A small DOCK and a moored boat where a building meets open water
* LAMP POSTS on ground cells, fences on edges
- Colour: the player picks a colour from a pastel palette before placing; each building keeps
its colour, roofs a darker tone of it. Neighbouring colours should look harmonious.
═══════════ INTERACTION ═══════════
- Left click: add a block (on water → new building at ground level; on a building → add a floor)
- Right click: remove a block, with the architecture re-resolving around it
- Hover: a translucent preview of the block to be added
- Orbit, zoom, pan camera with smooth momentum
- Undo/redo (50 steps)
- A palette of 10 colours; an eyedropper to pick an existing building's colour
- Changes animate: a new block grows up out of the water with a gentle bounce, and neighbouring
modules morph to their new shapes rather than popping
═══════════ THE WORLD ═══════════
- Stylized WATER with gentle waves, a soft foam line where it meets buildings, and reflections
- TIME OF DAY slider: morning haze, bright noon, golden hour, and night — at night windows light
up one by one with warm emissive glow and lamps turn on
- Gulls that circle and land on rooftops; small boats that drift between docks
- Soft shadows, ambient occlusion, a gentle tilt-shift depth-of-field for that miniature look
- Ambient soundscape (Web Audio): water lapping, distant gulls, and a soft chime pitched by
height each time you place a block — building a tower should sound like a little melody
═══════════ EXTRAS ═══════════
- New seed button (generates a fresh grid); a few preset starter towns
- Save/load towns to localStorage; share via a compressed URL hash
- Screenshot (canvas → PNG) and a "photo mode" that hides UI and slowly orbits
- A one-time gentle hint overlay, then nothing — no goals, no score, no menus in the way
═══════════ VERIFY ═══════════
□ The grid is all quads, organic, and relaxes into pleasant shapes with no degenerate cells
□ Modules deform cleanly to irregular quads with no gaps or cracks between neighbours
□ Roofs, arches, stairs, gardens, balconies, and docks all appear automatically in the right contexts
□ Removing a block correctly re-resolves every affected neighbour
□ Night mode lights windows and lamps
□ Placement animation is smooth and neighbours morph rather than pop
□ Save/load and URL sharing round-trip exactly
□ 60fps with a town of 200+ blocks
Output the complete single HTML file, no truncation. The irregular quad grid and the dual-grid
module selection with deformation are the critical systems — if modules crack at the seams or
the grid has pinched cells, the whole thing loses its magic.Demo 5: Skyfire, the Fireworks Show Designer (3/5)
The last app is a fireworks show designer that plays shells on a timeline and syncs them to a beat.
What worked
Pressing any button launches the show. You select the type of fireworks you want and click Launch Show, and it plays the whole launch sequence. The Show Cinema mode lets you drag and drop shells, and Free Play blows up a shell wherever you use it.
What went wrong
Again the layout is not responsive, and you can see it on smaller screens.
"This is not expected from a frontier model like this."
I gave it 3 out of 5.
Prompt to rebuild it
Build "Skyfire" — a fireworks show designer. Place shells on a timeline, sync them to a beat,
and launch the show over a city waterfront at night. ONE self-contained skyfire.html. Modern
three.js via the importmap above (UnrealBloomPass, EffectComposer allowed). Particle physics,
audio, and all visuals procedural.
═══════════ THE PYROTECHNICS (physically flavoured, visually spectacular) ═══════════
GPU particle system (instanced or Points with custom shaders) handling 50,000+ live particles.
Each STAR (glowing particle) has position, velocity, colour, burn time, and drag; gravity and
air drag act on everything, so bursts expand fast and then droop.
Implement these shell types, each with a distinct look:
PEONY (spherical burst, no trails) · CHRYSANTHEMUM (sphere with sparkling trails) ·
WILLOW (long golden trails that droop and hang) · PALM (few thick rising arms) ·
RING (a flat ring, randomly oriented) · CROSSETTE (stars that split into four mid-flight) ·
STROBE (stars that blink) · CRACKLE (stars that burst into tiny crackling sparks with a
snapping sound) · KAMURO (heavy glittering gold that lingers) · HEART (a heart-shaped burst) ·
MINE (ground-launched fountain spray) · COMET (a single bright rising star with a long tail)
Shells rise with a faint lift trail, then burst at their set height. Bursts can be multi-break
(a burst that bursts again).
COLOUR BY CHEMISTRY — a fun educational touch: the colour picker names the metal salt behind
each colour (strontium → red, barium → green, copper → blue, sodium → yellow, calcium → orange,
a strontium+copper mix → purple, magnesium/aluminium → white sparkle).
═══════════ THE SCENE ═══════════
- A city waterfront at night: a skyline silhouette with lit windows, a bridge, and calm water
- The WATER REFLECTS every burst (a planar reflection or screen-space approximation)
- SMOKE: each burst leaves a slowly drifting smoke cloud that catches the colour of LATER bursts
behind it — this is what makes a real show look rich rather than clean
- A light wind setting that pushes smoke and trails
- Burst light briefly illuminates the water, bridge, and smoke
- Bloom tuned so bursts glow and halo without the whole frame blowing out
- Camera presets: from the crowd on the shore, from a boat, from a high rooftop, and a free
camera. Gentle idle sway.
═══════════ SOUND (Web Audio, synthesized) ═══════════
- Launch thump, rising whistle for some shells, burst boom, crackle snaps, willow fizz
- REALISM DETAIL: the boom arrives AFTER the flash, delayed by distance ÷ speed of sound
(≈343 m/s) from the camera position — so the far rooftop camera hears it later than the boat.
This delay is the single most convincing detail in the build.
- A built-in procedural music track (choose from 3 moods: triumphant, dreamy, electronic) with a
known BPM, so the show can be choreographed to the beat
═══════════ THE SHOW DESIGNER (the core UX) ═══════════
- A horizontal TIMELINE at the bottom, with a beat grid from the music's BPM and a waveform
- Drag shells from a palette onto the timeline; they SNAP to beats (toggleable)
- Each placed cue has: shell type, colours (1–3), burst height, launch position along the
waterfront (a small top-down position strip), size, and count
- NOTE: shells must be launched BEFORE their beat — the editor places the BURST on the beat and
automatically computes the launch time from rise duration. That's the detail that makes it feel
professional.
- Multi-select, copy/paste, drag to move, and mirror a pattern across the waterfront
- Preview a single cue on click; play the whole show from any point; scrub
- QUICK PATTERNS: one-click sequences — a left-to-right sweep, a fan, a chase, a finale barrage
- Undo/redo
═══════════ MODES ═══════════
- DESIGN: the timeline editor
- SHOW: full screen, UI hidden, the show plays from the crowd camera with the music
- FREE PLAY: click anywhere in the sky to launch the selected shell instantly
- 3 built-in demo shows (one per music mood), each with a proper grand finale
═══════════ EXTRAS ═══════════
Save/load shows to localStorage; share a show as a compressed URL hash so anyone can open it and
watch; screenshot; a performance preset (Low/Medium/High) scaling particle counts and bloom.
═══════════ VERIFY ═══════════
□ All 12 shell types are visually distinct and recognizable
□ Gravity and drag give bursts the right expand-then-droop shape; willows genuinely hang
□ Crossettes split mid-flight; multi-break shells burst twice
□ Smoke lingers, drifts with wind, and is lit by later bursts
□ Bursts reflect in the water
□ The boom is delayed by distance and the delay changes between camera positions
□ Bursts land ON the beat because launch times are computed backwards from rise time
□ The demo shows play start to finish with a real finale
□ Shared URLs reproduce the show exactly
□ 60fps with 50,000 particles at High
Output the complete single HTML file, no truncation. The GPU particle system, the beat-aligned
launch timing, and the distance-delayed sound are the critical pieces.What My Five Tests Tell You About Gemini 4 Argon
Put the five scores side by side and the pattern is hard to miss.
Google's launch posts line up with the first half of this. The company says Argon brings major improvements in spatial reasoning, mathematical precision, and visual coding. Its demos are a pagoda in Blender, a racing game built from desk photos and a 3D wizard chess set. I did not find a Google page that claims responsive layouts or polished web page design. The Android developer post makes UI claims for Jetpack Compose, which is a different stack from web pages. So my 2 out of 5 results do not contradict anything Google promises for the web.
Two honest limits apply to my read.
Three Follow Up Tests I Would Run
If you want to check my read on your own account, I would try these before changing any opinion.
Gemini 4 Argon Benchmarks
Benchmarks are where Argon looks strongest on paper, and also where the details matter most. Here is how the numbers split.
Where Google's Own Table Shows Argon Ahead or Behind
Google published a table with Argon at its highest thinking level (pass@1) next to GPT 6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Rival scores in it come from public leaderboards, the rival labs' own reports and, for GraphWalks and PostTrainBench, Google's own runs, and Argon's own scores on several rows are Google's runs too. A third party review by AlphaCorp counts Argon ahead on 13 of 19 rows, tied on 1 and behind on 5.
| Benchmark | Argon | Best rival in Google's table | Read |
|---|---|---|---|
| Vals Index | 68.9 | Opus 5.5 at 67.0 | Argon ahead |
| DeepSWE v1.1 | 77.9 | Opus 5.5 at 74.2 | Argon ahead |
| GraphWalks, 256K to 1M tokens | 84.2 | Astra at 71.8 | Argon ahead |
| FrontierSWE v2 | 55.0 | Astra at 65.5 | Argon behind |
| Terminal Bench 4.0 | 57.4 | Opus 5.5 at 66.4 (Astra at 58.2) | Argon behind |
| PostTrainBench | 45.3 | Opus 5.5 at 49.3 | Argon behind |
| Terminal Bench Science 0.1 | 57.6 | Astra at 68.1 | Argon behind |
| OSWorld 2.0, offline subset | 69.2 | Astra at 72.6 | Argon behind |
The pattern in the losses is useful. Several of the rows where Argon trails (both Terminal Bench rows and OSWorld) test agents that work inside a shell or a desktop. The rows where it leads cover code repair, long context and broad task suites. One caution on Terminal Bench Science: Google computed Argon's 57.6 itself, and Vals lists a lower 44.29 on its own page, so I would treat that row as unsettled. Google's own evaluation PDF has the full list if you want to check a row I left out.
What Independent Leaderboards Say
Outside Google, the picture is more mixed. These ranks were read on 13 and 14 October, so they may have moved by the time you read this.
| Source | Result | Link |
|---|---|---|
| Vals Index | Argon first at 68.90%, ahead of Sonnet 5.5 at 67.04, Opus 5.5 at 66.97 and Fable 5.1 at 65.83 | Vals |
| Vals Vibe Code Bench | 91.91, second place behind Sonnet 5.5 | Vals |
| Artificial Analysis Intelligence Index | Argon 53, GPT 6 Astra 53, GPT 6.1 Sol 52, and Claude Opus 5.5 at max effort 58 | Artificial Analysis, The Next Web |
| AA Omniscience hallucination rate | Argon 15%, against 51% for Astra and 54% for Sol. Argon accuracy was 50%, against 63% for Astra | The Next Web |
| Artificial Analysis Terminal Bench 4 | Argon 57%, Astra 59%, Opus 5.5 (max) 60%, Sonnet 5.5 (max) 64% | Artificial Analysis |
| Arena Text | gemini-4-argon-high first at 1525, marked preliminary with 4,903 votes. Opus 5.5 high scored 1508 | Arena |
| Arena WebDev | Ninth at 1678, marked preliminary with 1,772 votes. GPT 6 Sol (max) is eighth at 1686 | Arena |
Two things stand out to me.
One more note on conflicts. Google's table gives Opus 5.5 a Terminal Bench 4.0 score of 66.4, while Artificial Analysis measured 60% for Opus 5.5 and 57% for Argon. Different evaluators use different setups, so compare models inside one evaluator, not across two.
Gemini 4 Argon Pricing
Argon is priced as a paid tier API model with a launch discount.
API Price Per Million Tokens
| Item | Through 30 November 2026 | From 1 December 2026 |
|---|---|---|
| Input | $2 | $4 |
| Output, including thinking tokens | $10 | $20 |
| Cached input | $0.10 | $0.20 |
| Batch or Flex, input and output | $1 and $5 | $2 and $10 |
| Priority, input and output | $3.60 and $18 | $7.20 and $36 |
Source: Gemini API pricing page. Output price includes thinking tokens, so a high thinking run costs more than the visible answer suggests. There is no free tier for the API.
Price Per Task Is What Counts
A low price per token does not guarantee a low bill. Artificial Analysis found that Argon wrote about 62K output tokens per task, against 27K for GPT 6 Astra. At the introductory price its test cost $1.99 per task, which becomes $3.98 at the standard price, while Astra cost $3.26. So the discount is doing real work here, and I would recheck your own cost after 1 December.
Hands on testers saw the same effect in other ways.
| Report | What it measured | Result |
|---|---|---|
| ComputingForGeeks | Total cost of nine calls (3 tasks, 3 runs each), Argon at the launch price | Argon default $0.045, Argon high $0.169, Opus 5.5 $0.254, GPT 6.1 Sol $0.062 |
| AIWIZ | Reported cost of 16 jobs run twice | Argon high $1.12, Argon medium $0.37, Sol $0.39, Opus 5.5 $1.92, Sonnet 5.5 $0.89 |
The ComputingForGeeks test is cheap at the default setting but did not deploy, which I cover below. The report adds that from 1 December the same nine default calls would cost about $0.09, which puts Sol's $0.06 ahead. In the AIWIZ test, Argon was priced at launch rates and the others at their own prices, and at December rates the Argon totals become $2.25 at high and $0.73 at medium. Cost per successful task is the number I would watch, not cost per call.
Cost Claims From X
People are posting their bills. These are user claims that I could not verify, so treat them as stories and not as data.
These four posts point in opposite directions. That alone tells you the bill depends heavily on the task and the thinking level.
Speed, Thinking Levels and Long Output
Thinking Levels at a Glance
Argon has three thinking levels. Google's advice is to run your evaluations at all three levels, then pick the lowest one that meets your quality bar.
| Level | What Google says it is for | What I would use it for |
|---|---|---|
| Low | Fewest tokens and lowest latency, though quality can drop on hard tasks | Rewrites, extraction, short answers |
| Medium (default) | Recommended for most tasks, including complex code and agentic use | Everyday builds and edits |
| High | Complex reasoning, hard math and science, and the most difficult code or agent tasks. Pairs well with long decode | Big refactors and tasks where a miss costs a lot |
The lowest level is not always a safe pick. In the AIWIZ test (16 jobs, each run twice), Argon at medium passed 28 of 32 runs, against 32 of 32 at high. It also did no thinking at all on 22 of its 32 medium calls. On a 99,000 token archive test, high answered 4 of 4 questions correctly in about two minutes, while medium answered 3 of 4 in about five seconds, with one calculation error.
Timings From Other Tests
| Source | Timing |
|---|---|
| ComputingForGeeks | Average latency of 5.5 seconds at default, 15.3 at high, 12.1 for Opus 5.5 and 18.7 for GPT 6.1 Sol |
| AIWIZ | Median short job of 19.2 seconds at high and 4.2 at medium |
| My racing game | About 15 to 20 minutes to generate a full game (from my video) |
| @Adidotdev | About 10 minutes and 1.5% of a weekly quota on Argon, against 15 minutes and 4% on Opus 5.5 (test credited to maxt3chno) |
| @appllamaio | Same prompt with /goal: Argon 38 minutes, Fable 5.1 two hours 10 minutes, Astra one hour 25 minutes |
I am not quoting a tokens per second figure. Artificial Analysis shows about 100 to 116 tokens per second on Google's own endpoints, while OpenRouter shows 43 to 60, and the two trackers measure different things.
On Hacker News, one user called the model pretty slow, and another said it was near frontier but showed infinite reasoning loops. A third described an 800K line code audit where each subagent took about ten minutes. Google itself lists occasional slowness and timeouts as a known limit.
Does Argon Really Have 1M Output?
Yes, but only through long decode. The context window is 1,048,576 tokens, and input, thinking and output all share it. Normal calls return up to 65,536 output tokens, and Google's model page lists up to 1,048,576 with long decode mode in the Interactions API. A page that promises 1M output in a single normal call overstates it.
background and store to true and ask for more than 65,536 output tokens.@bobbyhalljr_dev describes the client side route: past the per request limit, the response comes back with a continuation marker and a token, and you resend the request with that token to keep going.
How to Get Gemini 4 Argon
Where you can use Argon depends on your plan. This is the part where sources disagree most, so I went to Google's own posts.
| Where | Who gets it | Source |
|---|---|---|
| Gemini app and AI Mode in Search | Google AI Pro and Ultra subscribers | Google launch post |
| Antigravity | Ultra subscribers | Google launch post |
| Gemini API and Google AI Studio | Paid tier developers. Google says it is not on the Free tier or the Google AI Pro plan, and Ultra subscribers can try it in AI Studio | Pricing page |
| Google Cloud | Available through Google Cloud for enterprises | Google Cloud blog |
| GitHub Copilot | Pro+, Max, Business and Enterprise plans, with a gradual rollout | GitHub changelog |
| Fairwind program | Vetted partners only, with a separate version that has cyber capabilities | DeepMind |
A few practical notes.
What People Are Saying About Gemini 4 Argon on X
I read about 30 posts on X from 13 and 14 October, read only, with no likes or replies from me. These are user claims. I paraphrased them and linked each one so you can read the original and check the video or screenshot yourself.
Official Posts
3D, Games and Builds
| Who | What they tried | What they said | Link |
|---|---|---|---|
| @genevieve__h | Photos of a home office desk turned into a playable racing game with Blender MCP and Three.js | A how to thread, shown as a working game | Post |
| @rubenssoto_ai | Opus 5.5 against Argon on a cinematic 3D animation of a Tesla style electric motor | Same prompt, two videos side by side | Post |
| @universalmodhq | Argon against Muse Wonder on a game development prompt | Two different games from the same prompt | Post |
| @hug0_plat | A 3D physics test | Says Argon beat Muse Wonder | Post |
| @aimlapi | Argon against GPT 6 Astra in 3D | Says the quality is not there yet, though Argon was 5.6 times cheaper | Post |
| @shota7180 | A near future 3D game | Exploration, car races, puzzles and a DJ booth in one build | Post |
| @vikktorrrre | Recreating Google's launch video | Says Argon rebuilt it in Three.js and HTML in 45 minutes for about $20 | Post |
| @hqmank | A harder pelican riding a bicycle front end test | Says Argon got the bike geometry, pedaling, cap and sunglasses right, with small interactive touches | Post |
Coding and Engineering Claims
| Who | What they tried | What they said | Link |
|---|---|---|---|
| @gosrum (Japanese) | A full ts bench run with Argon in Cursor | Says it is not up to the Claude family, but wastes fewer tokens than the GPT family | Post |
| @Richardomuhuai (Chinese) | Optimizing vLLM for Gemma 4 E2B on an A100 | Says throughput rose from about 37.9K to 108K tokens per second with identical output, with pull requests sent upstream. Google's launch post makes the same claim | Post |
| @bobbyhalljr_dev | Long output through the API | Says past the per request limit you resend the request with a continuation token, and built a small loop for it | Post |
| @appllamaio | The same prompt with /goal on three models | Says Argon took 38 minutes, Fable 5.1 took 2 hours 10 minutes and Astra took 1 hour 25 minutes | Post |
| @Adidotdev | Argon against Opus 5.5, test credited to maxt3chno | Says Argon took about 10 minutes and 1.5% of a weekly quota, against 15 minutes and 4% for Opus | Post |
The vLLM claim is the biggest one in the list. Google's launch post makes the same claim, so it comes from Google as well as from this user. I have not seen independent confirmation, so I would wait for the upstream pull requests to land before trusting the numbers.
Quality Opinions
| Who | Take | Link |
|---|---|---|
| @LexnLin | A good model in many areas and much better at 3D than earlier models, but thinks Opus 5.5 can still make something better | Post |
| @AndrewOnXYZ | Better than expected, matches Opus in his view, and better on music | Post |
| @bridgemindai | A solid improvement, but nowhere near Opus 5.5, and disappointed that Google is still behind | Post |
| @prathamdby | Mixed front end results: some sections really good, some plain bad, and Opus 5.5 the clear winner | Post |
| @MichaelKaiserAI | Antigravity has Argon, but only for Ultra, and Pro users are still waiting | Post |
What the Posts Add Up To
Where Gemini 4 Argon Struggles
Every model has weak spots, and Argon's show up in the same places across my tests and other people's.
Layout, Spacing and Responsive Design
This is where I lost Argon three times. Arena WebDev ranks it ninth in a preliminary result, so eight models sit above it. Bloomberg, as relayed by implicator.ai, reports that Google staff doubt real world coding matches the benchmarks, and that anonymous employees said Argon struggles with some coding tasks, including front end design. Implicator did not see the Bloomberg story itself, and Google denies that Argon underperforms. Treat this as a signal about coding in general and not as proof of a layout flaw.
Infrastructure Tasks That Have to Deploy
ComputingForGeeks ran three infrastructure tasks (a hardened nginx deployment on k3s, a Postgres to S3 backup script and an OpenTofu VPC), three runs each.
| Model | Passed static checks | Passed the deployed k3s test |
|---|---|---|
| Argon (four thinking settings) | 9 of 9 at default and high | 0 of 12 |
| Claude Opus 5.5 | 9 of 9 | 3 of 3 |
| GPT 6.1 Sol | 9 of 9 | 3 of 3 |
The failure was specific. Argon ran nginx as user 101 without any writable volumes, so the container hit a permission error on its cache directory. The answer looked right on paper and failed in a cluster, though the report says all three fixes deployed once Argon saw the error.
There are fair caveats. Argon went through OpenRouter at four thinking settings, while Opus and Sol ran at default effort on their own APIs, so the comparison is not perfectly even. Still, a result that looks valid and does not run is a pattern worth testing for in your own stack.
Terminal and Desktop Agents
Argon trails on Terminal Bench 4.0, Terminal Bench Science, FrontierSWE v2 and the OSWorld 2.0 offline subset in Google's own table, and Artificial Analysis's Terminal Bench 4 run put it at 57%, behind Astra at 59%, Opus 5.5 at 60% and Sonnet 5.5 at 64%. If your work lives in a shell, test before you switch.
Slowness, Loops and Heavy Output
Hard Limits of the Model
Who Should Use Gemini 4 Argon
Here is my read, based on my five tests and the evidence above. It is my judgment, not a rule.
| If you are... | Use Argon? | Why |
|---|---|---|
| Building 3D scenes, games or simulations | Yes, try it first | Two 5 out of 5 scores from me, plus Simon Willison's Blender result and many X builds |
| Working with long documents or large codebases | Yes, at high thinking | 32 of 32 runs and a correct 99,000 token archive test at high, plus a strong long context row in Google's table |
| Doing research where wrong answers hurt | Worth testing | The 15% hallucination rate is well below Astra and Sol in the same test, though accuracy is lower too |
| Shipping responsive websites and UI | Test first, and add a layout pass | My 2 out of 5 results, Arena WebDev rank 9 (preliminary) and mixed X reports |
| Writing infrastructure code | Deploy before you trust it | The 0 of 12 deployed result from ComputingForGeeks |
| Running terminal or desktop agents | Test first | It trails on several agent rows in Google's table |
| Watching your budget | Track cost per task | Output volume is high, and the price doubles after 30 November |
| On a free plan | Not yet | The API has no free tier, and the Gemini app needs AI Pro or Ultra |
Gemini 4 Argon vs Claude Opus 5.5 vs GPT 6 Astra
Here is the closest thing to a fair side by side from the sources I trust. The label n/a means I did not capture that rival's score, and every row links its source.
| Measure | Gemini 4 Argon | Claude Opus 5.5 | GPT 6 Astra | Source |
|---|---|---|---|---|
| Vals Index | 68.9 | 67.0 | 63.1 | Google's table |
| Artificial Analysis Intelligence Index | 53 | 58 (max effort) | 53 | Artificial Analysis, The Next Web |
| DeepSWE v1.1 | 77.9 | 74.2 | 74.1 | Google's table |
| FrontierSWE v2 | 55.0 | 62.3 | 65.5 | Google's table |
| Terminal Bench 4.0 | 57.4 | 66.4 | 58.2 | Google's table |
| OSWorld 2.0, offline subset | 69.2 | n/a | 72.6 | Google's table |
| AA Omniscience hallucination rate | 15% | n/a | 51% | The Next Web |
| Arena Text score | 1525 (preliminary) | 1508 | n/a | Arena |
| Cost per task in Artificial Analysis's test | $1.99 intro, $3.98 standard | n/a | $3.26 | Artificial Analysis |
Argon is last of the four models in Google's Terminal Bench 4.0 row (Fable 5.1 scored 57.9). My read of the table:
No single model wins every row, which matches the advice ChatGPT gave me when I asked it: judge a model by cost per successful task on your own work.
More Prompts to Try With Gemini 4 Argon
The five app prompts are in the demo sections above. These four are new. I have not run any of them, so treat them as starting points and edit them for your project.
Responsive Audit Prompt (Untested)
Paste this after any front end build.
Audit the page you just built for responsive problems. Check these widths: 1440, 1024, 768 and 390 pixels. For each width, list: - any horizontal overflow - text that wraps badly - buttons or cards that overlap - cards or sections with no gap between them Then fix every problem in the code. Rules: - Use CSS grid or flexbox with gap values from an 8 point scale. - Do not use fixed pixel widths on containers. - Include the meta viewport tag. - Collapse a three column grid to two columns at 1024 pixels and one column at 768 pixels. Return the full updated file and a short list of what you changed.
Spacing Rules Prompt (Untested)
Add this block to the end of any UI prompt, since spacing was my biggest miss.
Spacing rules you must follow: - Use an 8 point spacing scale only (8, 16, 24, 32, 48, 64). - At least 24 pixels between a heading and the paragraph below it. - At least 24 pixels between a paragraph and its buttons. - At least 24 pixels between neighbouring cards. - Section padding of at least 64 pixels on desktop and 32 pixels on mobile. After you write the code, check every rule against it and fix any that fail.
Deploy Check Prompt for Infrastructure (Untested)
This one comes from the nginx failure in the ComputingForGeeks test.
Write the Kubernetes manifest for [your service]. Before you answer, check these points and show me the result of each: 1. Which user does the container run as, and can that user write to every path the process needs? 2. List every directory the process writes to and the volume that backs it. 3. Does the image work with the security context you chose? Then give me the final manifest and the commands to test it on a local cluster.
A Simple Scoring Rubric
My video scores come from watching each finished app. If you want a repeatable version, give one point for each line.
| Point | Question |
|---|---|
| 1 | Does the app run on the first try? |
| 2 | Is every feature from the brief present? |
| 3 | Does it look right at desktop width? |
| 4 | Does it look right at 768 and 390 pixels with no overflow? |
| 5 | Do the controls feel intuitive? |
This rubric is not exactly how I scored, so do not compare your totals with my 17 out of 25.
Frequently Asked Questions (FAQs)
How good is Gemini 4 Argon?
It is strong at 3D, long documents and code repair, and weaker at page layout and shell agents. It leads the Vals Index at 68.9% but scores 53 on the Artificial Analysis Intelligence Index, while Opus 5.5 scores 58. In my five app test it scored 17 out of 25.
What can Gemini 4 Argon do?
It reads text, images, video, audio and PDFs, and it writes text and code. Google highlights spatial reasoning, mathematical precision and visual coding. It has a context window of 1,048,576 tokens and it cannot generate images or audio.
Is Gemini 4 Argon free?
No. The Gemini API has no free tier for this model, Google says it is not on the Google AI Pro plan for API use, and the Gemini app needs a Google AI Pro or Ultra plan. API prices are $2 per million input tokens and $10 per million output tokens through 30 November, then $4 and $20.
How do I get Gemini 4 Argon?
Pick the surface that fits your plan. The Gemini app and AI Mode need AI Pro or Ultra, Antigravity needs Ultra, and the API needs a paid developer account (Google says a Pro plan does not include it, and Ultra subscribers can try it in AI Studio). GitHub Copilot also offers it on Pro+, Max, Business and Enterprise plans with a gradual rollout, and the API model ID is gemini-4-argon.
Is Gemini 4 Argon better than Claude Opus 5.5?
It depends on the task. Argon leads Opus 5.5 on the Vals Index and DeepSWE, but Opus leads on the Artificial Analysis index and on Terminal Bench 4.0, and several X users say Opus 5.5 still does better work. Test both on your own prompt and compare cost per successful task.
Is Gemini 4 Argon good for front end and UI work?
Not reliably yet, based on my tests. My perfume site got 2 out of 5, and Argon ranks ninth on Arena WebDev in a preliminary result. Give it exact spacing rules and breakpoints, and run a responsive audit prompt after the first build.
Final Thoughts
Gemini 4 Argon is a strong pick for 3D games, simulations and long document work, and a risky one for page layout and responsive design. My two 5 out of 5 builds and my three weaker ones tell the same story as the leaderboards and the X posts. Your next step is simple: copy the Slipstream prompt, run it, then paste the responsive audit prompt on a front end build and score both out of 5. If you want my prompts for more model tests, subscribe and join me on promptslove.com.





