Quick answer: Ox Alpha is a free, 1M-context reasoning model released on August 20, 2026 by an anonymous third-party provider and served through OpenRouter under the model ID stealth/ox-alpha. It accepts text, image, and video input, outputs text, supports tool calling and structured output, and is explicitly positioned for coding, sustained agentic work, and production workloads. During the preview it costs $0 for both prompt and completion tokens.
If you've watched the model feed this week, you know the pattern: a new model appears, the specs look borderline impossible for the price (free?), there are no benchmark scores, and the developer is anonymous. So you're left with the practical questions no teaser post answers: Is this real? Can I trust it? What is it actually good at — and is "free" going to disappear tomorrow? This article answers those questions using the OpenRouter official listing, public API data, and official announcements, updated August 22, 2026. We have not run our own benchmarks; everything below is based on official sources, with community claims clearly labeled.
Ox Alpha at a Glance
| Spec | Value |
|---|---|
| Model ID | stealth/ox-alpha |
| Release date | August 20, 2026 |
| Context window | 1,048,576 tokens (1M) |
| Max output | 131,072 tokens |
| Input | Text, image, video |
| Output | Text |
| Pricing (preview) | Free — $0 prompt / $0 completion |
| Provider | "Stealth" (anonymous third party) |
| Reasoning | Always on; default effort max (supports max / high / low) |
Those six numbers in the table are the whole story at a glance: a model with an enormous context window, a large output budget, multimodal input, and a free preview price tag. Everything else — what it's for, whether you can trust it, and who's already using it — is a question of how you read the official documentation.
What the Official Positioning Actually Says
The OpenRouter listing describes Ox Alpha in two sentences:
"Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads."
"It is suited for long-horizon software engineering, complex reasoning, and workflows that combine text with visual context."
Read them carefully: this model is not positioned as a general-purpose chatbot, a creative-writing model, or a cheap text-only utility. It's aimed at a specific job — software engineering and agent work that runs a long time on a complex task.
"Long-horizon software engineering" is the key phrase. A short-horizon task is a single question or a small edit: "explain this function," "fix this lint error." A long-horizon task is one that takes an agent many steps and lots of context to complete: reading an entire repository, planning a multi-file refactor, implementing it across dozens of files, running tests, iterating on failures, and reporting back. That's exactly what the 1M-token context window and 131K output budget are for — keeping the whole repo and the whole reasoning trace in context at once.
The API surface confirms the positioning. Reasoning is mandatory (reasoning.mandatory: true), with the default effort set to max and max / high / low supported. The model accepts tools, tool_choice, and response_format parameters, which means it works with agent harnesses and structured output pipelines, not just chat. Default temperature is 1 and top_p is 0.95. In other words: the wiring is built for agents, not for casual chatting.
The Trust Question Nobody Else Answers: How to Evaluate a Stealth Model
Here's the part most coverage skips. Ox Alpha is a "stealth model," and that fact creates a legitimate trust problem. Here's what "stealth" actually means on the record — and a concrete framework for deciding whether to use a model like this at all.
What stealth means. The OpenRouter listing states that Ox Alpha "is developed and operated by a third-party provider who has chosen to remain anonymous during this preview." The provider identity, architecture, and parameter count are not disclosed. Anonymity isn't a red flag by itself — anonymous previews are a known distribution pattern — but it changes what you can verify.
Who's responsible for what. OpenRouter's role is explicitly limited: it "routes requests to it and is not its developer, owner, or provider." OpenRouter is the delivery layer, not the maker. That matters for accountability: if something goes wrong, there is no public company behind the model to contact, and no history of releases to judge quality against.
The data policy, on the record. The listing states: "Prompts and completions are retained by the provider and are not used for training; all other use is governed by the Stealth Model Terms." So the provider retains your prompts and completions (retention, not training — a meaningful distinction), and the full policy lives in the Stealth Model Terms. If you're handling sensitive code or data, read those terms before routing production workloads through it.
What to evaluate when the developer is anonymous. With no brand reputation to lean on, check five things:
- Specs — Is the configuration coherent? A 1M context window with a 131K output budget, mandatory reasoning, and tool support is a consistent "agentic coding model" profile, not a random grab bag.
- Data policy — Retention without training is stated explicitly; verify the terms page for anything beyond that.
- Real usage signals — Raw token volumes and which tools are consuming them (more on that below) tell you whether actual builders are betting real workloads on it.
- Benchmarks — and their absence — There are no official benchmark scores on the model page, and the model is not listed on Artificial Analysis (checked August 22, 2026). The absence of numbers is itself information: treat "free frontier model" claims with caution until independent evaluation appears.
- The exit — Because it's a preview, everything — pricing, availability, and the provider's willingness to stay anonymous — can change without notice.
That framework is the honest way to use a model like this: adopt it for what it demonstrably provides (a free, 1M-context, multimodal reasoning model today), and keep an eye on the things it can't yet prove.
Is Ox Alpha a GLM Model? What the Community Is Guessing
You'll see this question everywhere, and the honest answer is: nobody knows, and anyone who states it as fact is guessing. The theory floating around the community is that Ox Alpha is a hidden multimodal variant of Zhipu AI's GLM family (such as GLM-5.3), based on comparisons of fingerprints, tokenizers, and video encoders. These are community-reported observations and have not been independently verified — they are speculation, not official information. No public statement from Zhipu AI, OpenRouter, or the anonymous provider confirms or denies any relationship. Treat every identity claim you read as unconfirmed until a named party says otherwise.
Why Is It Free?
The straightforward reading: Ox Alpha is in a preview period, and free access is the acquisition play. An anonymous provider gets distribution and real-world traffic in exchange for tokens, and the fastest way to build usage signals and mindshare in 2026 is a zero-price window. It's the same logic that put Ox Alpha on OpenCode Go's model list with "free for the next week" — near-unlimited usage that doesn't count against Go plan quotas.
What happens after the preview is genuinely unknown: post-preview pricing is not disclosed. The model page says "PRICE: Free" today, with no published roadmap for what comes next — so the practical move is trying it now while it costs nothing.
Who Is Actually Using It
The most concrete evidence that Ox Alpha is being taken seriously comes from the model page's own usage telemetry, covering roughly the first two days after release:
- ~657B prompt tokens and ~7.95B completion tokens processed
- Top 5 apps by traffic: Hermes Agent (120B tokens), Claude Code (108B), Oh-My-Pi (93.5B), DeepSeek Harness (multimodal-bridge) (84.8B), and ZCode (51.5B)
Read that list as a signal: every top consumer is an agentic coding or agent tool. The people burning through hundreds of billions of tokens aren't playing with it in a chat box — they're pointing autonomous coding agents at it. That's consistent with the official positioning (long-horizon engineering, sustained agentic work), and it's the strongest independent signal available that the model does something real in that niche. It is not a benchmark — raw demand doesn't prove output quality — but it's a meaningful real-world vote from agent-tool users.
How to Try Ox Alpha Yourself
You don't need an API key to poke at it. The fastest path is a browser: try Ox Alpha free on glm5.app — the page gives you a chat window pointed directly at stealth/ox-alpha, no setup required. If you're building, the model ID stealth/ox-alpha is what you'd pass to OpenRouter's API, and the API supports tool calling, structured output, and reasoning effort controls — so wiring it into an existing agent harness takes minutes. The OpenCode Go integration is another route: it lists Ox Alpha with a limited free window that doesn't count against your plan.
Whatever route you take, this is the moment to test it — while it's free and before the community speculation resolves.
Try It Before the Preview Window Closes
The preview is the best time to stress-test Ox Alpha: free tokens and no commitment beyond a chat box. Start a free Ox Alpha chat on glm5.app and throw a repo-scale task at it — a multi-file refactor, a bug that spans modules, or a long agent run that needs the whole codebase in context — before the free window closes and the identity rumors resolve into something more certain.
FAQ
Is Ox Alpha free? Yes — during the preview period, prompt and completion tokens are $0 on OpenRouter (the page lists "PRICE: Free"). It also appears on OpenCode Go with a limited free window. Post-preview pricing is not disclosed.
Who made Ox Alpha? An anonymous third-party provider that has chosen to remain unidentified during the preview. OpenRouter states it only routes requests and is not the developer, owner, or provider. The provider's identity, architecture, and parameter count are not disclosed.
Is Ox Alpha a GLM model? Unconfirmed. Community speculation links it to Zhipu AI's GLM family based on fingerprint and tokenizer comparisons, but those claims are not independently verified, and no official party has confirmed or denied a connection.
What is the context window of Ox Alpha? 1,048,576 tokens (1M), with a maximum output of 131,072 tokens.
Does Ox Alpha accept images and video? Yes — it accepts text, image, and video input and outputs text, which suits workflows that combine visual context with reasoning.
Where can I try Ox Alpha?
The quickest way is a browser chat on glm5.app, pointed at stealth/ox-alpha. You can also call the model directly through OpenRouter's API (model ID stealth/ox-alpha) or use it via OpenCode Go during its free window.
Sources
Last updated: August 22, 2026. Based on the OpenRouter official listing, public API data, and official announcements. We have not run independent benchmarks; community-reported figures are labeled as such and not independently verified. The provider is anonymous and the preview free window may change — check the official pages for current terms and pricing.
