Ox Alpha has been live for less than 48 hours, and it's already the most-tested free model on OpenRouter this week. Between the public Sulat demo gallery and a steady stream of builder posts on X, a real picture is forming of what people are actually doing with it — beyond the "mystery model" headlines. Here are the ten most notable, sourced examples, including the ones that don't hold up as well as their viral framing suggests.
Update — August 26, 2026 (evening): Z.ai named the model GLM-5.3-Flash — MIT weights, 320B-A18B, $0.15/$0.50 API. Official launch guide.
For the full spec sheet and setup guide, see explainx.ai's Ox Alpha / GLM-5.3-Flash overview; for the identity story — forensics through Bloomberg to the named SKU — see full evidence timeline.
Update — August 26, 2026: Z.AI confirmed Ox Alpha is a GLM model; open weights promised that night. Ox Alpha hit #1 on OpenRouter, more than doubling DeepSeek usage.
TL;DR
| # | Use case | Verdict |
|---|---|---|
| 1 | GPU-accelerated fluid simulation | Visually stunning — but likely reproduces an existing open-source reference |
| 2 | DeepSWE coding benchmark | 80% self-reported, ahead of Fable and GPT-5.6 Sol |
| 3 | Cloned Ubuntu desktop UI | One-shot desktop-environment recreation |
| 4 | 3D Snakes and Ladders | Full 3D board game, playable |
| 5 | 3D swordfighting demo | Combat simulation in 3D space |
| 6 | 3D cherry blossom scene | Interactive 3D visual |
| 7 | Branded landing pages | Movie/franchise promo pages |
| 8 | Typewriter UI simulation | Interactive text-input experience |
| 9 | Agent harness backend | Billions of tokens via Claude Code, Hermes Agent |
| 10 | Multimodal video/image tasks | 1M context, native image + video input |
1. A GPU-accelerated fluid simulation — the viral one
Developer scaling01 posted that Ox Alpha one-shot a GPU-accelerated fluid simulation in a single 1,000-line HTML file, calling it "1000x better" than what Qwen3.8 27B or Opus 4.5 produced on the same prompt earlier that week. The clip genuinely looks good — smooth, colorful, physically plausible fluid dynamics rendered directly in a browser.
Video from scaling01's post on X, re-hosted on explainx.ai for playback.
But the replies caught something the original post didn't mention: developers Sebastian Macke and Nitor both pointed out the output closely matches paveldogreat's WebGL Fluid Simulation — a well-known, publicly available open-source project that's been online for years. The likely explanation isn't that Ox Alpha solved real-time fluid dynamics from scratch; it's that a model with strong code retrieval reproduced a reference implementation it had almost certainly seen during training. That's still a genuinely useful capability — reliably reproducing a known-good implementation on request is not nothing — but it's a different claim than "novel one-shot physics engineering," which is how the demo was originally framed.
2. An 80% score on the DeepSWE coding benchmark
Developer Ben Davis ran Ox Alpha through DeepSWE, a coding-agent benchmark, and reported it scoring 80% — ahead of Fable at 65% and GPT-5.6 Sol at 52%. This is the closest thing to a formal capability number in circulation, though it's still a single developer's self-reported eval run rather than a published, audited leaderboard entry. It lines up directionally with the traffic data on OpenRouter, where Claude Code and Hermes Agent alone push billions of tokens through the model.
3. A one-shot Ubuntu desktop clone
Among the demos catalogued at ox-alpha.demos.sulat.com is a recreation of the Ubuntu desktop environment — window chrome, taskbar, and the general visual language of a real Linux desktop, generated in a single pass. Desktop-environment clones are a recurring stress test for frontier coding models because they require getting dozens of small UI details right simultaneously, not just one flashy centerpiece effect.
4. A playable 3D Snakes and Ladders
The same gallery includes a full 3D reinterpretation of Snakes and Ladders — the classic board game rebuilt with three-dimensional pieces and board, with working game logic rather than just a static render. Board-game recreations are a good proxy for whether a model can juggle game-state logic and visual presentation in the same generation pass.
5. A 3D swordfighting combat demo
Also in the Sulat gallery: a 3D swordfighting simulation, with character models and combat interaction rendered in three-dimensional space. This is a heavier lift than a static 3D scene — it implies animation state, collision handling, and interactive controls generated together.
6. An interactive 3D cherry blossom scene
A more purely visual entry: an interactive 3D cherry blossom scene, showcasing rendering and particle-style effects rather than gameplay logic. Demos like this are a useful signal for anyone evaluating the model for creative/visual coding work rather than application logic.
7. Branded landing pages
Two entries in the demo gallery are branded promotional landing pages — a "Primetime" movie promo page and a Spider-Man-branded page — testing the model's ability to produce polished, on-brand marketing web design rather than functional application code. This is a common real-world use case for teams doing rapid landing-page prototyping.
8. A typewriter UI simulation
A smaller but telling demo: an interactive typewriter experience, simulating the visual and interaction feel of typing on a physical typewriter in the browser. Small, self-contained interactive toys like this are a fast way to spot-check a model's attention to animation timing and micro-interaction polish.
9. Backend for real agent harnesses, at scale
Beyond one-shot demos, Ox Alpha is already running as the actual backend for production agent workloads. OpenRouter's own app-level traffic data shows Claude Code (~9.3B tokens) and Hermes Agent (~9.0B tokens) as the top senders on the model, alongside Oh-My-Pi, DeepSeek Harness, and Z Code. That's a meaningfully different signal than a flashy demo — it means teams are routing real, sustained coding-agent work through it, not just testing it once.
10. Multimodal tasks combining text, image, and video
Ox Alpha's 1,048,576-token context window and native image and video input support open a use case most of the flashy one-shot demos don't touch: workflows that mix long text context with visual inputs — a repo plus screenshots, a design mock plus a spec document, or a video walkthrough plus a bug report. explainx.ai's full spec breakdown covers the exact context, output, and throughput numbers if you're evaluating this specifically.
Try it yourself
The model shipped as GLM-5.3-Flash on August 26, 2026 — MIT weights on Hugging Face, API at $0.15/$0.50 per M tokens. Route to z-ai/glm-5.3-flash when OpenRouter lists it, or use Z.ai's API. The free stealth/ox-alpha preview may deprecate without notice. Full launch details: GLM-5.3-Flash official post. Given how much of the current buzz rests on demos that look good at a glance, the most useful thing you can do with any of these examples is run your own version of the prompt rather than taking a viral clip's framing at face value.
Related on explainx.ai
- GLM-5.3-Flash official launch — Ox Alpha unmasked
- OpenRouter Ox Alpha: full specs, pricing, and setup
- Ox Alpha: forensics timeline
- Hermes Agent #1 on OpenRouter rankings
- Choosing open-weight vs. closed AI models
- AI token pricing, explained
Primary sources: scaling01 on X · Ox Alpha demo gallery, Sulat · paveldogreat's WebGL Fluid Simulation · Ox Alpha on OpenRouter
Demos sourced from public builder posts during Ox Alpha stealth week. Official product: GLM-5.3-Flash (MIT). Updated August 26, 2026.