Gemini 4 Argon for web design

Gemini 4 Argon is Google's frontier model, announced on September 30, 2026 for a group of cyber defenders and generally available since October 13 on the paid Gemini API as gemini-4-argon. It is also in Google AI Studio, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode for AI Pro and Ultra subscribers.

On Design Arena at the time of writing, Argon is tenth on the Website board at 1315 and third on the Image to HTML board at 1269. It ranks higher when it is handed a picture than when it is handed a sentence, and that is the most useful fact about it for anyone building a site.

The short call:

• A full page in one pass? Argon, with its 1M token output limit and an average of under four minutes per site on the Website board.
• Rebuilding a reference or a mockup? Argon is among the top three on that board.
• One section against a written spec? gemini-3.8-flash, which averages 139 seconds per site on the same board and is enough for a single block.
• A page that stops looking generated? That comes from the specification, and Thanor writes one per design.

Google pitches Argon on visual work: 3D visualizations, games and interactive animations. These three pages are that kind of work, each built from a prompt that decided everything first:

What Gemini 4 Argon is#

Google launched Argon in two steps. The September 30 post introduced it as a model for real-world coding, enterprise knowledge work and cyber defense, rolling out first through the Fairwind Program. The October 13 post opened it to developers, enterprises and subscribers.

The context, output and input figures are from Google's models page in AI Studio; the output increase and the prices are from Google's launch posts.

The output limit is the number that changes a web build. At 64K tokens a long single-file page with inline styles, scripts and copy could run out before the footer. At 1M, a whole site fits in one response with a wide margin.

The Gemini 3 write-up warned against planning around a Gemini generation beyond the 3 series, because none had shipped when it was written. Argon is that generation. Read that piece for the Flash advice, which still holds, and this one for the flagship. A second Gemini 4 model, Carbon, exists but is not broadly available; it is covered further down.

Where Argon ranks on the design boards#

Design Arena figures were read on November 9, 2026, and Arena's WebDev board carries a November 8 date. Argon is not yet on Design Arena's agent-built Web Apps board, so that row is missing.

The Website rank rests on 10,978 battles, more than any model in the top ten except Muse Spark 1.3, so it is one of the best-measured numbers on the board. Argon sits 10 points under Claude Opus 5.5 there, which means Opus is expected to win about 51 votes in 100 against it, and 85 points above Gemini 3.1 Pro Preview at 1230.

Speed is Argon's quiet advantage. It averages 235 seconds per site on the Website board, against 309 for Claude Opus 5.5, 836 for Kimi K3 and 1,087 for GPT-6 Astra at max effort. For a designer who regenerates a hero ten times, that gap compounds.

On Image to HTML, where the input is a screenshot or mockup, Argon climbs from tenth to third, behind GPT-6 Astra and Meta's Muse Ember and ahead of Kimi K3 and Claude Opus 5.5. The tasks differ, so this is not proof of cause. It does fit how Google describes the model, and it suggests a way to work with it. For the cross-model view of all these boards, read Design Arena, explained.

What Google says Argon does for visual work#

Google's October 13 post claims "major improvements in spatial reasoning, mathematical precision, and visual coding". The examples it gives are the clearest guide to where Google expects Argon to shine:

• Googlers using it to prototype complex 3D visualizations, high-fidelity games and interactive animations, in vanilla code that renders directly in the browser.
• A chess-style 3D game wired for the web in Three.js with GLSL shaders and shatter physics, the kind of build collected in Three.js website examples.
• A detailed pagoda model built with multi-agent visual critique loops in Antigravity.
• A Blender workflow driven through Blender MCP from a single generated image.

Two things are missing from those posts, and both matter here. Neither launch post makes a claim about web design, typography or interface layout as such, and neither mentions a design leaderboard. The claims are about 3D, games, animation and code quality. The benchmarks Google cites are mostly engineering, agent and science benchmarks, such as 77.9% on DeepSWE v1.1.

Read that as a strength in mechanics, not taste. A model that is good at spatial reasoning and shader code will build the scroll-scrubbed film or the 3D object you describe. It will not decide that the object should be emerald and gold on a green-black ground, or that the reveal should land on cubic-bezier(0.22, 1, 0.36, 1) over 700ms. Those choices arrive with the brief or they arrive as defaults, and the defaults are the same ones every model ships.

Google's own design tool points the same way. Stitch, its UI design app in Google Labs, open-sourced its DESIGN.md format in April so that design rules can travel between tools. A format for writing the decisions down is an admission that the decisions have to be written down. Claude Design vs Google Stitch compares the two design tools directly.

Hand Argon a picture and a page of values#

Argon's best board takes an image as input. You can give it the same advantage on your own page: decide the design, then hand it over as both a picture and numbers.

Attach the reference#

Argon accepts images and video. Attach a frame of the hero you want, a mood still or the poster image of a film the page will play. A picture carries composition, scale and color relationships that a sentence cannot.

Write the values next to it#

A picture leaves the numbers open, so state them. A fragment written for this article:

palette   ground #0A1410, text #EFEBE0, accent #2FBF8F on one element per section
type      display 112px at 0.88, body 17px at 1.6, ratio 1.28, two weights
motion    every entrance cubic-bezier(0.22, 1, 0.36, 1) over 700ms
scroll    one film, 180 frames, scrubbed over 3 screens, poster until it paints
order     film stage, collection rail, services, atelier, footer
refuse    no gradient behind text, no third typeface, no icon set

Give it the media#

A model cannot invent a film of your product or a cut-out of your object, and a page that needs one will get a placeholder or a stock substitute. Thanor prompts carry hosted films, cut-outs and stills with exact URLs, and the public animated backgrounds page offers looping clips whose URL you can paste straight into a brief. For how to keep a film behind text fast and legible, see video backgrounds done right. These two pages run on a looping film each:

State what it must refuse#

Three or four refusals hold a page's character better than any adjective. The Gemini 3 brief shows the pattern, and it carries over to Argon unchanged.

Use the long output carefully#

A 1M token output limit removes the old reason to split a page into many requests. It does not remove the reason a long page drifts: the further a section sits from the brief, the more the model leans on its own defaults.

Three habits keep a long Argon build on the brief:

1. Restate the token block at the end of the request as well as the top. It costs about sixty words and keeps the footer obeying the header.
2. Keep the brief in a file, not in the chat. A file can be diffed, versioned and pasted into a different model next month, which matters when the next launch arrives.
3. Split by job, not by length. One request for the whole page in Argon, then single sections in gemini-3.8-flash against the same values, pasted fresh with every request.

Google's model list is short enough to plan around. At the time of writing:

Carbon is worth knowing about because it shows where Google is heading. Google's October 28 post says it gained 89 points to reach 1768 on WebDev Arena, where it now sits third and is marked preliminary. Until it opens up, it is a benchmark entry rather than a tool.

The verdict on Gemini 4 Argon#

Gemini 4 Argon is a top-ten design model by public vote, faster on the Website board than the Anthropic, OpenAI and Moonshot models around it, and a top-three model when it starts from a picture. Its output limit makes a full single-file site in one pass practical.

Ranked by effect on the page you get:

• The brief and the reference: values plus an image decide the look.
• The media you supply: a real film or cut-out separates a designed hero from a stock one.
• Argon over Flash: holding a long build together.
• Gemini over another maker: speed, price and whether you work in Google's tools.

Argon is the Gemini to build with today. Carbon may change that when it opens, and the arena numbers keep moving in the meantime, which is one more reason to keep the brief in a file that outlives the model id. The same page of values goes into Argon, into Flash and into whatever ships next.

For the head-to-head with Anthropic's lineup, read Claude vs Gemini for web design, and for how the leading models compare overall, the pillar on choosing a model.

Give Argon something to rebuild#

The fastest test of Argon is the one its best board runs: a page whose design is already decided. Open Vesperion and copy its prompt; it is free with an account. Paste it into AI Studio with gemini-4-argon selected, and compare the page that comes back with the moving preview.

Every value in that prompt is a decision Argon no longer has to guess, and its media URLs give it the real film to play. If a value comes back changed, the fault is easy to find, because you know exactly what was asked.

When the result holds up, take the next design from the library and run the same check on a page you would sell. Argon is fast enough on that board to run the comparison twice in the time some models take once.