Ox Alpha in one sentence
Ox Alpha is an anonymous "stealth" AI model built for coding, sustained agentic work, and complex reasoning, with a 1M-token context window. It appeared without an announced vendor — first on OpenRouter around August 20, 2026, then as a free limited-time model inside OpenCode — and quickly climbed the trending lists because it's both strong at code and free to use.
Where it came from (and who's really behind it)
Stealth models are usually a way for a lab to A/B a new release under a codename before attaching its brand. Ox Alpha followed that pattern exactly: no official vendor at launch, a fresh model ID on OpenRouter, and a burst of "what is this thing" posts. The community quickly guessed it was a Zhipu / GLM-family model from its behavior, benchmark shape, and routing — and that guess turned out to be right.
On August 26, 2026, Z.AI (Zhipu) confirmed to Bloomberg that Ox Alpha is a new GLM-series model, and the official name landed the same evening: GLM-5.3-Flash — a 320B-A18B (mixture-of-experts) model released under an MIT license with open weights. During its four-day free preview, Ox Alpha climbed to #1 on OpenRouter, more than doubling DeepSeek's usage, which is a rare thing for a brand-new stealth model. So the identity is no longer a mystery: it's GLM-5.3-Flash, and you can now run it either through hosted APIs or the open weights themselves.
Ox Alpha at a glance
| Property | Value |
|---|---|
| Real identity | GLM-5.3-Flash (Z.AI / Zhipu) — confirmed Aug 26, 2026 |
| Model ID (preview) | stealth/ox-alpha on OpenRouter |
| Architecture | 320B-A18B (mixture-of-experts), MIT license, open weights |
| Context / max output | 1,048,576 tokens in · 131,072 tokens out |
| Modalities | Text, image, video in → text out |
| Tool calling | Supported |
| Throughput / latency | ~50 tok/s · ~2s P50 (OpenRouter dashboard, preview) |
| Launched | OpenRouter (Aug 20, 2026), then OpenCode as free limited-time |
| Strengths | Coding, agentic loops, front-end UI generation, long-context |
| Price now | ~$0.15 / $0.50 per 1M input/output tokens (free preview ended) |
Is Ox Alpha actually good?
The early reports are genuinely positive for a free model. It scores well on coding benchmarks (DeepSWE-style tasks in the ~80% range in some tests), handles long files thanks to the huge context window, and gets singled out for clean front-end / UI generation. For everyday coding and agent work it's very usable. As always with stealth models, benchmarks are early and can shift once the model is finalized and priced.
Ox Alpha vs. today's frontier models — coding tasks solved
% of tasks solved on an independent community benchmark of 10 real-world coding tasks (as published on the official oxalpha.com, 2026). Reference models scored as passes out of 4 attempts per task; Ox Alpha (GLM-5.3-Flash) solved 8 of 10 — ahead of every frontier model it was lined up against, including one task no reference model could reliably pass. Small third-party sample: directional, not definitive. Benchmarks shift monthly; treat any single ranking as a snapshot.
Context window vs. typical models
A 1M-token window means Ox Alpha can hold roughly an entire mid-size codebase (or a very long agent run) in context at once — the headline reason it trended for agentic work.
How to run Ox Alpha in a coding agent
Ox Alpha is exposed through OpenAI-compatible routes (OpenRouter / OpenCode-style endpoints), so any agent that lets you set a base URL, a key, and a model name can drive it. The friction with a lot of tools is that they're welded to one provider — to try a new model you're editing env vars or waiting for official support. The smoother path is an agent where the provider, base URL, key, and model are just settings you pick from a list.
Point an agent at Ox Alpha
- Get an endpoint + keyGrab an OpenAI-compatible endpoint that serves Ox Alpha (e.g. via OpenRouter) and an API key.
- Set base URL + keyIn your agent, set the base URL to that endpoint and paste the key. No translating proxy needed for OpenAI-compatible providers.
- Select the modelSet the model identifier the provider exposes for Ox Alpha, then run your normal coding tasks.
- Keep secrets outBecause it's a free stealth model that may log prompts, don't feed it anything sensitive.
The cleaner way to try any stealth model
If your goal is "I want to test Ox Alpha today and switch to something else tomorrow without re-plumbing my setup," use an agent built to bring your own model. Wuwei is a free, open-source (MIT) AI coding agent that runs locally: you pick the provider, base URL, key, and model from a list, so trying a new stealth model — or falling back to Claude / GPT / a local model — is a click, not a config rewrite. Your code stays on your machine, which matters even more when the model on the other end might be logging.
FAQ
What is Ox Alpha?
Ox Alpha is an anonymous "stealth" AI model tuned for coding and agentic work, with a 1M-token context window. It launched free for a limited time on OpenRouter in August 2026 and later on OpenCode.
Is Ox Alpha GLM-5.3?
Yes — on August 26, 2026, Z.AI (Zhipu) confirmed to Bloomberg that Ox Alpha is GLM-5.3-Flash, a 320B-A18B mixture-of-experts model released under an MIT license with open weights. It launched anonymously as the stealth model "ox-alpha" on OpenRouter, hit #1 during its free preview, and was then revealed.
Is Ox Alpha free?
It was free during its limited-time preview on OpenRouter and OpenCode (Aug 2026), which retained prompts — so people were told not to send secrets. Now that it's revealed as GLM-5.3-Flash, the API lists at roughly $0.15 / $0.50 per million input/output tokens, and the open weights (MIT) can be self-hosted for free.
How do I use Ox Alpha in a coding agent?
It's served over OpenAI-compatible endpoints, so any agent that lets you set a base URL, key, and model can run it. Wuwei is a free, open-source agent where you pick the provider and model from a list, so you can point it at Ox Alpha (or any model) without editing env vars.