Dev48
Language
  • About
  • Services
  • Industries
  • Technologies
  • Articles
  • Contacts
Book a call
    Home/Articles/Gemini 4 neon video an artifact export not a generation
Dev48
info@dev48.org+375 29 320-40-96WhatsApp@isky_by

© 2026 · All rights reserved.

Gemini 4 Neon Video: An Artifact Export, Not a Generation

Source: OrcaRouter•November 22, 2026 at 12:02 AM

Gemini 4 Neon Video: An Artifact Export, Not a Generation

A 60-second Gemini 4 Neon film isn't generated video — Google's card lists text as the only output. What the skill really does, and what it costs.

At 08:07 UTC on 21 November 2026, the account @dotey — 宝玉 on X — posted a 60.03-second clip at 1920×1080 with a single line of Chinese above it: 做视频效果也不错, roughly "it's not bad at video either." The subject was G​emini 4 Neon, G​oogle's week-old model, and the link under the sentence pointed at an open-source project called open-artifact, described in the post as the video-production Artifact (skill). The clip had 230 likes and 76 replies when we read it on 22 November. It looks like a produced piece of motion design.

It is not a generated video. G​oogle's own model documentation for G​emini 4 Neon lists one output type and one only — text — and the film that came out of it is a program: HTML, CSS and JavaScript scenes drawn as a function of time, plus a synthesised score, rendered by a local web page and exported to MP4 by that page's export button. The distinction is not pedantry. It changes what the model is good at, what a second cut costs, and whether the thing you are buying is a renderer or a writer.

What the post actually says

The whole public record of the claim is a sentence tag and a 60-second attachment. There is no prompt, no scene count, no thinking level, no token count, and no statement of whether the film came out of the API, G​oogle AI Studio, Antigravity or something else. The post does not even say the clip was made with the skill it links — it presents the skill as the way to do this kind of work, with the clip as the reason to care. Read it as a recommendation with a demo attached, not a test.

One thing is worth knowing before weighing the recommendation. The repository is five days old: GitHub records it as created on 17 November 2026, with 135 stars and 25 forks at the time of writing, Apache-2.0 licensed, and a release — 0.12.2 — cut on 21 November, the same day as the post. Its README says it began life as baoyu-design, and the credit line in the licence section reads "Copyright 2026 Jim Liu 宝玉". That is the same 宝玉 the posting account uses as its display name. So the person telling you Neon is good at this is showing a tool that carries their name on the copyright line. That does not make the demo false. It does make it a maker's demo, and the honest reading is that it demonstrates what the pipeline can produce in skilled hands, not what an average prompt produces on the first try.

The model card says output: text

G​oogle shipped G​emini 4 Neon on 18 November 2026 at $3 per million input tokens and $15 per million output, with cached input at $0.15. The launch post describes it reading spreadsheets, PDFs, charts, audio and video. The model page goes further and settles the question the post leaves open: under Supported data types it lists "Inputs: Text, Image, Video, Audio, and PDF" and then, on its own line, "Output: Text." The same page marks audio generation as not supported. Context is 1,048,576 tokens, with 65,536 tokens of output by default and up to the full million with long decode.

That asymmetry is where the confusion starts. Video is something this model watches, not something it makes. The card view on the same documentation site is looser — it prints "Modalities: Text, Image, Audio, Video, PDF" without saying which direction is which — and a reader skimming that line could reasonably conclude Neon outputs video. It doesn't. In G​oogle's own stack, generating frames is a different model line; Neon's contribution to a moving picture is the code that draws it.

How a text model produces a film

The motion artifact type inside open-artifact is explicit about the mechanism. A film is a folder of small JSON documents: a scene list with stage size and frame rate, one JavaScript module per scene that draws it from a time variable t, a score module that synthesises music and effects, and an optional narration module for captions and voiceover. Nothing is rendered as video by the model. The local page plays the documents the way a browser plays a page, and Export turns that playback into a file.

The skill's own documentation is candid about the split of labour, and the details are the ones that matter if you are deciding whether to try this:

• Length — a film runs 10 to 30 seconds unless you ask for more, with a hard ceiling of 300 seconds. The clip in the post is 60.03 seconds, so its author asked.

• Export — MP4 at 720p by default or 1080p, at the full, half or a quarter of the frame rate, plus GIF, HTML and ZIP. The attachment is 1920×1080, the non-default option.

• Who writes the file — the instructions tell the agent to make no video or audio file at all. Export is the page's job, and stills for checking a scene are drawn by headless Chrome.

• Safety rails — no scene under a second, no more than three cuts or flashes per second, and any captions come from lines the person supplied rather than lines the model invents.

So three things happen in sequence: the model writes a small program, a browser draws it, and an encoder — the project ships Rust/WebAssembly crates for GIF and MOV encoding — writes the container. Only the first step involves G​emini 4 Neon. A 60-second 1080p film with a score is a serious amount of code for one model to get right in a handful of turns, and that is a real compliment. It is a compliment about software engineering, which is exactly where Neon's published numbers are strongest.

What one film costs on Neon's price list

Nothing in the post discloses token usage, so the following is arithmetic on G​oogle's list price rather than a measurement. A film of ten scenes means a long first turn, several read-and-write rounds against the document store, and a fair amount of rereading as the model checks its own scenes. Call it 150,000 input tokens and 25,000 output tokens for a working film: $0.45 of input at $3 per million and $0.375 of output at $15 per million, or about 83 cents. A leaner pass at 40,000 in and 8,000 out comes to 24 cents. Cached input is where the number really moves — at $0.15 per million, 120,000 cached input tokens cost about two cents instead of 36, and the whole run drops under 50 cents.

Set that against what independent testing costs the model per task. Artificial Analysis scores G​emini 4 Neon at 61 on its Intelligence Index at xhigh effort, at 106 output tokens per second, 22.3 minutes and $6.44 per index task. A film is cheaper than an index task, which is the useful surprise here: the expensive part of this workflow is not the film, it is the iteration.

And iteration is where the economics invert. A per-second video model bills every take; a code-drawn film bills the tokens that write it and nothing for the render. Change a colour, re-export, and you have spent zero tokens — the cost is your own CPU and however long headless Chrome takes to draw 1,800 frames. That is a genuinely different cost curve from generative video, and it is the strongest argument in the post's favour that the post itself doesn't make.

Worth stating plainly: G​emini 4 Neon is not in our catalogue. The two G​emini 4 models OrcaRouter does serve are Gemini 4 Argon and Gemini 4 Carbon, both at $2.00 in and $10.00 out per million at provider list price with no markup on top — which is the comparison a reader actually needs, because Carbon is the model Neon is priced against and Argon is the cheaper way to run the same artifact loop while you decide. Both sit behind one key, so a code-writing session that times out mid-film can fail over without a second contract or a code change.

Where the claim is weakest

The strongest version of the sceptical case is not that the clip is fake — the pipeline is public and the mechanism is documented, so anyone can reproduce the shape of it. It is that a good code-drawn film is a coding benchmark wearing a director's hat, and this model's coding numbers are already its best numbers. Neon sits at 58.6 on the Coding Agent Index and 65.2% on Terminal-Bench 4.1 by G​oogle's own reporting, so "good at making films" may be mostly "good at writing animation code," which we already knew and which does not generalise to a model that writes worse code.

Three specific gaps remain after reading everything available:

• Attribution — no settings are disclosed, and the 60-second runtime and 1080p resolution are both departures from the skill's defaults, so somebody made choices we cannot see.

• Selection — one clip is a highlight, and a maker's demo is selected by the maker. There is no failure rate, no first-attempt example, and no comparison against the same prompt on Carbon.

• What "good" means — the visual quality partly belongs to the template and the skill's design rules, not to the model. A different artifact type with weaker scaffolding would flatter the model less.

None of that is disqualifying, and it cuts both ways: a code-drawn film is also auditable in a way a diffusion render is not. If a frame is wrong, it is wrong in a file you can read and a line you can point at.

How to try it yourself

The path is short enough to describe in a paragraph. The project needs Node 22.18 or later, ships a plugin bundle through GitHub Releases, and lists Antigravity among its supported hosts — install the extracted folder with agy plugin install, or copy it into ~/.gemini/config/plugins/open-artifact, and restart. That happens to be the host where G​emini 4 Neon is the default model for new conversations on plans that include it, so the two pieces meet without configuration. The first prompt is a sentence: a twenty-second launch animation for this app, or a title card for that project. Then you look at stills before you spend anything on an export, which is the workflow the skill itself recommends.

Do that once before believing the post, and watch two numbers rather than the picture: how many turns it takes before the first scene draws correctly, and how much of the input bill is cache reads. Those tell you whether the loop is cheap for your work. The clip tells you it is possible; only your own token count tells you whether it is worth it.

The open question is not whether G​emini 4 Neon can end up in a video file — it demonstrably can, through a program it wrote. It is whether G​oogle treats that as a first-class output. Right now the artifact layer is third-party, the export is a local page's job, and the model's documented output is text. If that changes, the honest framing of this clip will change with it: from a clever use of a coding model to a feature listing. Until then, the compliment worth repeating is the narrower one — Neon writes animation code well enough that someone who builds the tooling chose to show it off with it.

← All articles

What this article says

More in AI & Machine Learning

All →
Data Processing Incidents on Inference Providers
OpenRouter

Data Processing Incidents on Inference Providers

Orchestrator: One API for frontier and open-weight models
Perplexity

Orchestrator: One API for frontier and open-weight models

Gemini 4 Hydrogen: What xhigh Buys for 6x the Bill
OrcaRouter

Gemini 4 Hydrogen: What xhigh Buys for 6x the Bill

Security event: Prompt leakage for requests in June and July of 2026
Together AI

Security event: Prompt leakage for requests in June and July of 2026

Qwen4 Max is live on Together AI
Together AI

Qwen4 Max is live on Together AI

How we secure 100+ clusters across 20+ clouds
Baseten

How we secure 100+ clusters across 20+ clouds

More from OrcaRouter

Gemini 4 Hydrogen: What xhigh Buys for 6x the Bill
OrcaRouter

Gemini 4 Hydrogen: What xhigh Buys for 6x the Bill

Gemini 4 Neon vs Claude Haiku 5.5: The Ends of One Ladder
OrcaRouter

Gemini 4 Neon vs Claude Haiku 5.5: The Ends of One Ladder

Hy4 flash Vision: 74.8 MMMU-Pro, and Why 52.0 Still Stands
OrcaRouter

Hy4 flash Vision: 74.8 MMMU-Pro, and Why 52.0 Still Stands

FIBO 2.0 Is on Civitai Now: Read the Commercial-Use Line Twice
OrcaRouter

FIBO 2.0 Is on Civitai Now: Read the Commercial-Use Line Twice

FIBO 2.0 vs HY-Image 3.5: One Board Calls It a Tie
OrcaRouter

FIBO 2.0 vs HY-Image 3.5: One Board Calls It a Tie

GigaChat3.5-Omni Leak: Sber's 432B Model Grew Eyes and Ears in a Draft PR
OrcaRouter

GigaChat3.5-Omni Leak: Sber's 432B Model Grew Eyes and Ears in a Draft PR