A new AI model appeared on OpenRouter and OpenCode on August 20, 2026, and nobody knows who built it. It goes by the codename Ox Alpha, it accepts text, images, and video, it reads up to a million tokens at once, and for roughly one week it costs nothing to use. Within a day of launch, coding agents had pushed billions of tokens through it, and Reddit turned into a detective forum, with users dissecting its tokenizer, its error messages, and even the language of its backend logs to figure out which lab is hiding behind the release.
Here is everything we know so far: the specifications, the early test numbers and what they actually prove, what Reddit users are reporting, the leading identity theories, and exactly how to try the model before the free window closes.
What Is Ox Alpha?
Ox Alpha is what the industry calls a stealth model: a frontier-class AI system released anonymously through an API platform so the lab behind it can collect real-world feedback before a public launch. The official OpenRouter listing describes it as “a reasoning model designed for coding, sustained agentic work, and production workloads.”
In plain terms, it is built for long jobs: multi-step coding, autonomous agent runs, and reasoning across huge amounts of material in one session.
No name, no company, no press release
The model is listed under the provider name “stealth,” and OpenRouter states plainly that it only routes the requests. The system was developed and operated by a third party that has chosen to remain anonymous during the preview. As of August 21, no company has claimed it.
The fifth mystery model in six months
This playbook is now familiar. Ox Alpha is the fifth anonymous release in about six months, and the previous four all followed the same script: a quiet debut, a burst of free traffic, then a company stepping forward. All four turned out to come from Chinese labs, including Zhipu AI’s GLM-5, Xiaomi’s MiMo-V2-Pro, Ant Group’s Ling-2.6-flash, and Meituan’s LongCat-2.0. That track record shapes most of the current speculation.
Ox Alpha Specifications at a Glance
| Specification | Detail |
|---|---|
| Context window | 1,048,576 tokens |
| Maximum output | 131,072 tokens |
| Inputs | Text, images, video |
| Output | Text only |
| Tool support | Function calling, structured JSON output |
| Pricing | Free input and output tokens during the preview |
| Throughput | 28 tokens per second (median) |
| Latency | About 4 seconds (median) |
| Uptime | 99.99 percent over the first days |
| Capacity claim | 100 trillion tokens per day during the free week |
Two details stand out. Video input is a first for a stealth release, and only a handful of named frontier models handle it well. And the one-million-token context window could, in theory, hold an entire mid-sized codebase or months of documents in a single request. One caveat: a big context window does not automatically mean a model uses every token effectively, and nobody has independently verified how Ox Alpha performs on extremely long inputs yet.
Does Ox Alpha Really Beat GPT-5.6 Sol and Claude Fable?
The claim spreading fastest across social media is a benchmark result: Ox Alpha reportedly scored above 80 percent on DeepSWE software engineering tasks, against roughly 65 percent for Claude Fable and 52 percent for GPT-5.6 Sol.
Those numbers need context. The result came from a user-run test containing just 10 tasks. The official DeepSWE benchmark contains 113 long-horizon engineering tasks written from scratch across 91 active open-source repositories in five programming languages, specifically designed so that reference solutions never leak into training data. A hand-picked sample of 10 can favor one model purely through task selection.
The accurate conclusion is narrower: Ox Alpha performed impressively in one small preliminary test, and that is all. At the time of writing, the model has no published results on any major independent leaderboard.
What real usage suggests
Adoption tells its own story. Coding agents such as Claude Code and Hermes Agent harnesses pushed a combined 18 billion tokens through Ox Alpha almost immediately, which means real developers are treating it as a daily workhorse rather than a curiosity. For a brand-new anonymous deployment, the platform metrics above are solid production numbers.
What Reddit Users Are Saying
Early reactions are genuinely mixed, which makes them more useful than the hype.
The praise
Some developers report strong results on exactly the work the model claims to target. One user said Ox Alpha found two genuine bugs in a Python project that other auditing tools had missed. Another asked it to build an interactive weather application: the model initially spent several minutes reasoning and then replied without doing the work, but after a second prompt it produced an app the tester judged more accurate and less buggy than a version from DeepSeek V4 Flash. OpenCode users describe it as effective at building frontend interfaces, making basic backend fixes, working from screenshots, planning larger coding tasks, and burning fewer tokens than some other free models.
The complaints
The negative reports matter just as much. Users say the model can reason for minutes without taking any action, struggles with complex backend projects, and performs inconsistently when building from scratch. In one comparison it failed a problem that GPT-5.6 Sol solved cleanly. Others find its prose flat and mechanical and still prefer DeepSeek for everyday writing.
The economics question
Reddit is also asking the obvious question: serving a frontier-class model for free at this scale is expensive, so what does the anonymous provider get in return? The realistic answer is data, feedback, and attention. Testers are effectively running the lab’s evaluation program for free. That trade can still be worth taking, as long as you understand you are part of it.
Who Built It? The Leading Theories
There is no confirmed answer, but the community investigation has produced several candidates.
| Theory | Supporting evidence |
|---|---|
| Multimodal GLM-5.3 variant (Z.ai) | Tokenizer matches the GLM family, Chinese-language backend messages, API error code 1210 previously associated with Z.ai, similar speed and caching patterns |
| Xiaomi MiMo V3 | Xiaomi has run anonymous previews before, and the 100-trillion-token capacity claim resembles earlier MiMo promotions |
| Tencent Hy4 | Proposed in a dedicated Reddit investigation thread |
| MiniMax | Circumstantial timing, an unannounced model would fit |
| A router, switching between several models | Would explain inconsistent behavior across task types |
The GLM theory currently leads. Testers also noticed Ox Alpha trips over the same rare token glitches that affect GLM, MiMo, and Qwen models, which again points to a Chinese lab. There is one wrinkle: the publicly available GLM-5.3 is text-only, while Ox Alpha accepts images and video, so if the theory holds, this would be an unreleased multimodal variant. Until the developer steps forward, all of this remains unofficial fingerprinting.
How to Try Ox Alpha Yourself, Free
There are three ways in, depending on how technical you want to get.
Option 1: Chat on OpenRouter, no code needed
- Open the Ox Alpha model page on OpenRouter.
- Select Try this model or Playground.
- Sign in with a free OpenRouter account.
- Submit a task as text, an image, or a video clip.
This is the best route if you just want to feel the model out.
Option 2: Call it through the API
Developers can use OpenRouter’s OpenAI-compatible API with the model ID stealth/ox-alpha. Generate an API key in the OpenRouter dashboard, point any standard SDK at the OpenRouter base URL, and set the model field accordingly. Both prompt and completion tokens are billed at zero during the preview.
Option 3: Use it inside OpenCode
The launch is a joint promotion with OpenCode, the open-source coding agent, which is offering the model for about a week with rate limits generous enough that the team calls it near unlimited usage. Connect the OpenCode Zen provider, open the model catalog with the /models command, and select Ox Alpha.
Whichever route you pick, a real but non-sensitive task reveals far more than a chat prompt: fix a bug in a small project, add a feature plus its tests, or hand the model a screenshot of an interface to analyze. These test planning, tool use, and the ability to finish long work.
Read the Fine Print Before You Paste Anything Sensitive
Free frontier compute always comes with conditions, and this launch has a wrinkle worth understanding. OpenCode’s announcement promotes zero data retention, yet the OpenRouter listing states that the anonymous provider does retain prompts and completions, just without using them for training. The applicable policy may depend on how you access the model, and until the developer publishes clear terms, the safe assumption is that your inputs are stored by a company that has not told you its name.
The practical rules: do not submit passwords, personal information, confidential documents, or proprietary source code. Remember also that the free period has no published end date and could change overnight, and that no audited benchmarks exist yet, so every performance claim rests on small community samples.
Why Stealth Releases Keep Happening
Anonymous launches solve real problems for AI labs. They generate unbiased feedback, since testers cannot bring brand expectations to a model with no brand. They stress-test infrastructure with real traffic before an official launch. And they build organic attention: a mystery model that beats named competitors in community tests earns weeks of free coverage no ad campaign could buy. Expect the pattern to continue, which means evaluating an unknown model quickly is becoming a genuinely useful skill.
The Coursiv App and Platform Experience
Launches like this one are exactly why learning AI fundamentals pays off. New models now appear monthly, and the people who benefit are the ones who can open an unfamiliar tool, probe its strengths in an afternoon, and fold the good ones into their workflow. That skill is learnable. Coursiv teaches it through short, structured daily lessons covering practical prompting, comparing and choosing AI tools, and applying them to everyday work tasks, all designed for busy people without a technical background. The app guides you step by step, so when the next Ox Alpha drops, you will know exactly how to put it through its paces. As with any learning product, check the official site for current course details and pricing.
What to Watch Next
If history repeats, the lab behind Ox Alpha will identify itself within days or weeks, most likely alongside an official launch with a real name, published benchmarks, and pricing. Whether the fingerprint sleuths guessed right about GLM will be the first thing everyone checks. The second question is cost: previous stealth models converted their free previews into aggressively cheap public pricing, which is part of how Chinese labs have pressured Western competitors all year. A frontier-level coder at a fraction of frontier prices would ripple far beyond this one model.
The bottom line: Ox Alpha is a rare chance to test a possible frontier model at zero cost, with the small print that you are part of the experiment. Try it on a non-critical project this week; just do not build a production workflow on an anonymous preview.