On August 20, 2026, a model with no maker appeared on OpenRouter under the ID stealth/ox-alpha. Three days later, OpenCode’s live data page showed roughly 12 trillion tokens processed, 180,000 unique users, and 3.56 million sessions, making it the #2 model there by recent usage. That is the fastest zero-to-everywhere run of any stealth model this year, and it happened without a single press release, because there is no one to issue it.
Here is what is actually confirmed, what is theory dressed as fact, and what to do about it. This is a living page: when the maker is revealed or the pricing lands, we update this same URL the same day.
On this page · 6 sectionsOpen
- Ox Alpha is an anonymous stealth reasoning model that appeared on OpenRouter on August 20, 2026 as stealth/ox-alpha: 1,048,576-token context, 131K max output, text, image, and video input, tool calling, and $0 preview pricing.
- Adoption exploded in three days: OpenCode’s live data page showed roughly 12 trillion tokens, 180,000 unique users, and 3.56 million sessions by August 23, making it the #2 model there by recent usage.
- Nobody has claimed it. The leading community theory is Z.ai’s GLM family, backed by a 44-of-44 tokenizer fingerprint match one researcher published August 23, but a Gemini theory circulates too, and self-identification by the model proves nothing.
- The viral 80 percent coding score is 8 of 10 hand-picked tasks, not a full benchmark; community hands-on reports are sharply mixed, from repository-scale wins to repeated identical mistakes.
- The free window is expected to be short, with community estimates around August 27 to 28, unconfirmed, and the retention terms differ between access routes: OpenCode says zero retention, OpenRouter’s listing says the provider retains prompts.
- This is a living page: when the maker is revealed or pricing lands, we update the same URL the same day.
- What is Ox Alpha?
- An anonymous preview reasoning model on OpenRouter (stealth/ox-alpha), live since August 20, 2026: 1M-token context, multimodal input, tool calling, free during the stealth window. The maker is undisclosed.
- Who made Ox Alpha?
- Unknown. The leading theory is Z.ai’s GLM family, based on published tokenizer fingerprinting; a Gemini theory also circulates. No lab has claimed it, and we update this page when one does.
- Is Ox Alpha free?
- Yes, for now: $0 input and output on OpenRouter, and a limited-time free tier inside OpenCode. Community estimates put the window’s end around August 27 to 28, unconfirmed.
- Is it actually good?
- Mixed early record. The 1M context appears real (needle tests pass to ~934K tokens), long agent sessions draw praise, but users also report 503s, stalls, and repeated coding mistakes. The viral 80 percent score is 10 hand-picked tasks.
§ 01What is confirmed
The facts below come from OpenRouter’s live model listing and OpenCode’s documentation, as of August 23, 2026:
| Item | Status |
|---|---|
| Model ID | stealth/ox-alpha on OpenRouter |
| First appeared | August 20, 2026 |
| Developer | Undisclosed; OpenRouter states it only routes for an anonymous provider |
| Context window | 1,048,576 tokens |
| Max output | 131,072 tokens |
| Inputs | Text, images, video |
| Agent features | Tool calling, structured output, reasoning-effort settings |
| Preview price | $0 input, $0 output |
| Weights | Not released; no self-hosting |
The 1M context appears to be real rather than a metadata placeholder: one researcher’s needle-retrieval tests passed at placements around 934K tokens, failing only above roughly 1.005M.
§ 02The identity mystery, honestly
Nobody has claimed Ox Alpha, so everything about its origin is inference. The inference is getting good, though. The leading theory is Z.ai’s GLM family: an August 23 investigation ran about 600 calls and 44 discriminating tokenizer strings and reported a 44-of-44 match with the GLM-5-generation tokenizer. Chinese-language validation errors have leaked through the API, and a separate analysis argues its video-token behavior matches GLM-5V. A competing theory says it is an unreleased Gemini checkpoint, fueled by a DeepMind researcher’s hints on X.
Two things are worth keeping straight. First, tokenizer fingerprints are strong evidence of a model family, not an official attribution. Second, the model naming itself in chat proves nothing: models reproduce labels from training data all the time. The honest summary is “probably GLM-lineage, officially unknown,” and we will replace this whole section the day someone claims it.
§ 03Is it actually good?
The viral number says 80 percent. The fine print says that was 8 of 10 hand-picked coding tasks in one independent test, not a full benchmark run, and broader community testing lands closer to “strong but not obviously frontier-beating.”
The genuinely impressive reports cluster around long-horizon work: a user migrated two large projects that previously forced constant context compaction in other tools, with the model finding and fixing bugs across a 100K-to-200K-token working set. Users who put it in a real agent harness with tools and a terminal consistently rate it higher than users who just chat with it.
The complaints cluster just as clearly: 503 errors and dying streams under the load spike, sessions stalling until manually nudged, one user watching it apologize for a mistake and immediately repeat it, and overthinking at high reasoning-effort settings. Standard stealth-preview behavior, in other words: interesting model, overloaded infrastructure, no guarantees.
§ 04The part worth thinking about before you build on it
Ox Alpha is free, anonymous, and temporary. Each of those matters:
Free means the window is the product. Community estimates put the end around August 27 to 28, and while that specific date is unconfirmed, every previous stealth codename followed the same arc: vanish, then reappear under a real name with real pricing. We track the money side, including what each access route costs after the window, in our pricing page.
Anonymous means the data question has no good answer yet. The retention statements even disagree by route: OpenCode says the provider follows zero retention, while OpenRouter’s own stealth-model listing says prompts and completions are retained by the provider. Until the operator has a name, sending it credentials, customer data, or sensitive proprietary code is a bet on an unknown counterparty.
Temporary means anything you wire against the free endpoint should assume the endpoint dies. That is not a criticism; it is what previews are for.
§ 05What Ox Alpha is not
It is a model, not an agent, and the distinction is where all the practical decisions live. Ox Alpha accepts inputs and emits text or tool calls. Everything that makes that useful for real work, the tools themselves, filesystem and browser access, permissions, retries, verification, memory that survives the session, comes from the harness around it. That is why its heaviest early usage is inside coding harnesses rather than chat interfaces, and it is the same reason we argued in our harness ranking that the harness layer, not the model layer, is where working AI actually gets decided.
That layer is also where we live, so discount accordingly: CellCog is the employee layer above models like this one. Models rotate under our platform as they leapfrog each other, and the point of a standing AI employee is that the rotation does not matter to you. The employee keeps its role, its context, and its continuity; the model underneath is an implementation detail that gets better every few months, sometimes via a mystery box on OpenRouter.
§ 06What happens next
Three outcomes are live. The reveal: some lab claims it, pricing appears, and the free-window users become launch-day customers, which is very likely the strategy. The quiet death: the endpoint vanishes and the model reappears renamed, the standard stealth arc. Or the long preview: it stays free and anonymous longer than expected, soaking up training-adjacent feedback.
We update this page the same day any of them happens.
Q1What is Ox Alpha and where did it come from?
Ox Alpha is a stealth-preview AI model that appeared on OpenRouter on August 20, 2026 under the ID stealth/ox-alpha, with no named developer. OpenRouter states it only routes access for an anonymous third-party provider. It spread into OpenCode and other coding-agent clients within a day.
Q2Is Ox Alpha GLM or Gemini?
Officially, nobody knows. The strongest published evidence points to Z.ai’s GLM family: one August 23 investigation reported a 44-of-44 match against the GLM-5-generation tokenizer across ~600 calls, and Chinese-language error strings have leaked through the API. A competing Gemini theory exists. A model naming itself in chat is not evidence either way.
Q3What are Ox Alpha's specs?
Per OpenRouter’s live listing: 1,048,576-token context, 131,072-token max output, text, image, and video input, text output, mandatory reasoning with effort settings, tool calling, and structured output. Weights are not public and there is no self-hosting option.
Q4Can I trust Ox Alpha with my code?
Treat it as an anonymous provider, because it is one. The retention statements even disagree by route: OpenCode says zero retention, while OpenRouter’s stealth-model listing says the provider retains prompts and completions. Do not send credentials, customer data, or sensitive proprietary code until the operator is named.
Q5How do I try Ox Alpha?
Through OpenRouter’s API with model ID stealth/ox-alpha (OpenAI-compatible clients work), or through OpenCode’s limited-time free tier. It is a preview: rate limits, availability, and pricing can change without notice.
Q6What happens when the stealth window ends?
Typically the model disappears and reappears later under its real name with real pricing, as earlier stealth codenames did. Nothing built on the free endpoint should assume it survives. We will update this page the day the reveal happens.