A ghost shipped a frontier model.
On August 20, 2026, a model called Ox Alpha showed up on OpenRouter under the ID stealth/ox-alpha, along with availability through OpenCode. No launch post from a known lab. No founder thread. No pricing page. Just a listing with a 1,048,576-token context window, free access, and a claimed allowance of up to 100 trillion tokens per day for coders.
I curate agent tooling for a living, which means I spend most of my week separating real capability from marketing copy. Ox Alpha inverts that job entirely. There is no marketing copy. There is no company. There is a model endpoint and a lot of developers finding out what it does by pointing their agents at it.
What is actually confirmed
Keeping this tight, because the confirmed set is small and the speculation pile is enormous:
- Ox Alpha is an anonymous model released in August 2026 with no known developer.
- Context window of 1,048,576 tokens, commonly rounded to 1M.
- Free access, with a claimed ceiling of up to 100 trillion free tokens per day for coding work.
- Accepts text, image, and video inputs.
- Positioned for sustained software engineering, complex reasoning, and coding tasks.
- Distributed via OpenRouter and OpenCode.
That is the list. Everything beyond it — which lab is behind it, what the training data looks like, whether the free tier survives the month — is guesswork. The name itself is part of the ambiguity. People type it as 0x Model because “Ox” reads like a hex prefix, which tells you something about who found it first.
Why the context number matters for agents specifically
A million tokens sounds like a spec-sheet flex until you run an agent loop. Then it becomes the difference between an agent that can hold a repository in working memory and one that keeps rediscovering the same files.
Most agent frameworks spend a surprising share of their engineering effort on context management: chunking, retrieval, summarizing prior turns, deciding what to drop when the window fills. That machinery exists because the window was small. When the window gets large enough to hold a mid-sized codebase plus the full history of a long task, you can delete a lot of that plumbing and let the model see everything at once.
The “sustained software engineering” framing points the same direction. Long-running tasks are where agents fail most often, and they usually fail because state got lost between steps, not because the model could not write the function. Multimodal input adds another lane here — text, image, and video means an agent can take a screen recording of a bug or a screenshot of a broken layout as part of the same conversation, without a separate vision pipeline bolted on.
The free-tier question you should think about before you plug it in
A claimed 100 trillion tokens per day at no cost is not a business model. It is either a stealth evaluation run, a distribution play, or a benchmark of how much load developers will generate when the meter is off. In all three cases, the generosity is a phase, not a guarantee.
My practical read for anyone building on it:
- Treat it as an experiment, not infrastructure. Free anonymous endpoints do not come with uptime commitments.
- Keep your model layer swappable. If your agent talks to OpenRouter through an abstraction, swapping
stealth/ox-alphaout later costs you a config change instead of a rewrite. - Be deliberate about what you send. An anonymous model with no known operator means no published data handling policy you can point a security reviewer at. Proprietary code and customer data are the wrong test payload.
- Log your prompts and outputs locally. If the endpoint disappears, your evaluation data should not disappear with it.
What this release says about the space
Stealth listings on aggregators are becoming a normal way to test a model in public. It is a clean feedback loop: ship to an audience that will stress-test coding performance within hours, collect real usage patterns, skip the reputational risk of a named launch that underdelivers.
What is new is the scale of the freebie and the size of the window arriving together. That combination invites exactly the workload a lab would want data on — long agent runs against real repositories, generating exactly the traces that make the next version better at the same job.
So use it, and use it with clear eyes. Run it against a task you already have a baseline for, ideally something long and multi-file where the context window should show its value. Compare it to whatever model your agent uses today on the same task, with the same scaffolding. Write down what you find.
Anonymous or not, a model that can hold your whole project in its head is worth an afternoon of testing. Just do not build your roadmap on an endpoint that nobody has claimed yet.
🕒 Published: