AI

ox-alpha: The Free Stealth Model on OpenRouter Is GLM-5.3-Flash

Dark cover: the codename ox-alpha beside a glowing blue fingerprint dissolving into particles - the anonymous stealth model traced by its tokenizer

On 20 August 2026 a frontier-class AI model appeared on OpenRouter with no company name attached to it. It is called ox-alpha, it is free for about a week, it reads text, images and video - and at first nobody who built it would admit to it. We pointed it at a real production task the day it dropped. Here is what it is, what it can do, and who is behind it.

Update, 26 August 2026: Z.ai has named it. In the official GLM-5.3-Flash launch post, the lab wrote that it tested the model anonymously as ox-alpha on OpenCode and OpenRouter before the release. The fingerprints below had been pointing at the right family - the product name is GLM-5.3-Flash, a cheaper sibling of GLM-5.3, not GLM-5.3 itself. Weights are on Hugging Face under an MIT license.

What ox-alpha actually is

OpenRouter lists it under the provider name “Stealth” and calls it a reasoning model for coding, sustained agentic work and production workloads. In plain terms: a capable, general-purpose model aimed squarely at software engineering.

$0free on OpenRouter during the preview window
1.05Mtoken context window, up to 131K output
text · image · videomultimodal input, tuned for coding and agents

It went live on 20 August 2026, and the OpenCode team ran a parallel promotion putting it at zero cost for the following week, with rate limits generous enough that they called it “near-unlimited.” The unusual part is the data policy: the provider states that prompts and completions are not used for training - OpenCode calls it zero data retention - which is rare for a stealth release. You can reach it right now through the OpenRouter API or playground, or through OpenCode in a terminal, under the model ID stealth/ox-alpha.

Try it while it’s free: the model lives at openrouter.ai/stealth/ox-alpha. The free window opened on 20 August 2026 and was set to run about a week; pricing after that has not been announced.

So who made it?

When we first published, nobody had confirmed anything. Two labs kept coming up - and one of them had the receipts.

Z.ai / GLM-5.3 - the fingerprint case

A developer known as dax ran a tokenizer test across 25 prompts; ox-alpha’s token counts matched Z.ai’s GLM-5.3 exactly, off by a constant 75-token wrapper each time. A separate analysis matched ox-alpha’s full OpenRouter API contract - seven fields at once - to GLM-5.3 alone out of more than 400 models. Z.ai has been pushing GLM toward vision, which fits ox-alpha’s image and video support.

Xiaomi / MiMo - the precedent case

Xiaomi’s MiMo team has done exactly this before: MiMo-V2-Pro ran on OpenRouter as “Hunter Alpha” before Xiaomi claimed it at launch. ox-alpha’s spec sheet - a million-token context, native multimodality, an obvious lean toward agentic coding - matches the shape of what Xiaomi has been shipping.

Fact versus interpretation

The tokenizer and API-contract matches are real, reproducible measurements, and they prove shared lineage - almost certainly the GLM 5 generation. They did not, by themselves, prove ox-alpha was literally GLM-5.3. After this article first ran, Z.ai confirmed the model as GLM-5.3-Flash. One more clue had already pointed the same way: ox-alpha trips on the same “dirty token” that has previously snagged Qwen and GLM-family models.

We tested it the day it dropped

Speculation is cheap, so we ran ox-alpha against real work. We handed it - and Anthropic’s Claude Fable 5 - the identical brief for a secure, multi-tenant incident-intake app: specification first, then implementation, then a security review. Afterwards we attacked what each one built.

ox-alpha was fast and startlingly cheap: it finished in about an hour on roughly a fifth of the tokens, and produced a correct, readable application. But its isolation between customer organizations was a convention rather than a mechanism - one forgotten predicate would leak silently, a gap its own security notes admit - it wrote no tests, and its interface needed a manual reload after every change. Fable did the slower, more expensive, more thorough job. On a security build, that difference is the whole game.

See the full head-to-head: the exact prompt, the metrics, the screenshots and the scores are on our secure multi-tenant build benchmark, and everything we know about the model - including the identity evidence - sits on the ox-alpha model page.

Why a free stealth model shows up at all

ox-alpha did not appear in a vacuum. Over the past year the share of tokens processed by US models on OpenRouter has fallen from around 70% to roughly 30%, with Chinese labs - DeepSeek, Tencent, Xiaomi, Z.ai - taking the rest. A free, high-context, multimodal stealth drop is exactly the move that accelerates that shift: developers get a frontier-class model at no cost, and whichever lab is behind it collects a week of real-world usage data before it has to attach a price or a name.

Quick answers

Is ox-alpha free?

Yes, during the preview window that opened on 20 August 2026 - OpenCode put it at $0 for about a week with generous limits. Pricing after the window has not been announced.

What is ox-alpha?

The OpenRouter stealth preview of Z.ai’s GLM-5.3-Flash: a reasoning model built for coding and agentic work, with a ~1M-token context and text, image and video input.

Is ox-alpha GLM-5.3?

Yes - specifically GLM-5.3-Flash, confirmed by Z.ai on 26 August 2026. The tokenizer and API fingerprints already pointed at that family; the official name is the flash-tier sibling of GLM-5.3, not GLM-5.3 itself.

Is it safe to send it real prompts?

The provider states prompts and completions are not used for training. As with any third-party endpoint, don’t send secrets - but by the stated policy this one is unusually clean for a stealth release.

Weighing a new model for real work? That is most of what we do in AI development - we put frontier models through actual production tasks before we trust them with client code. If you are deciding what to build on, let’s talk.

Which workflow should you connect to AI?

Start with a clear task, define the data and permissions, and plan how to check the result.

You can also start with a short description of the workflow or product you want to build.

Ideas, delivered to your inbox

Occasional practical guides and examples on AI and building products. No spam.

Unsubscribe anytime.

Ask DevShift