Unsigned

Everybody is looking for the owner.

Ben Davis chooses ten “bugs.”

He does not invent them. They come from real open source projects. They are real errors. The model will read the code, find the problem, write the fix, run the tests, and try again if the fix does not work. So it will do the work of a developer.

First he tries the models we all know. GPT-5.6 Sol solves five of the ten bugs. Claude Fable 5 solves six. Then comes the turn of a model that appeared on OpenRouter one day before, with no announcement and with only a code name.

Eight.

He shares the result and writes under it: “I am very confused.”

His confusion makes sense. We do not know who made the thing he just tested. There is no technical report and no price. It sits on OpenRouter’s list without a real name: stealth/ox-alpha.

How it happened

20 August 2026. A new line appeared on OpenRouter’s model list. The model has a one million token context window, so it is at the highest level today. It can read text, images and video. The price field says zero. This is how stealth models are introduced now. And Ox Alpha is not the first one. On the same day, OpenCode, a coding tool, also announced free access to the model for one week. With a capacity of 100 trillion tokens per day.

The real story is the speed of the spread. OpenRouter’s own activity panel shows 12 billion tokens on the first day. Two days later, 4.4 trillion. The numbers on the OpenCode side are similar: about 12 trillion processed tokens, 180 thousand users, and the second most used model on the platform. And there was no press release and no launch event.

Is this possible?

This is the first question, and the answer is not as complex as people think. An empty GPU is like an empty seat on a plane that is taking off. An hour you do not sell never comes back. And usage rates are really not what people expect. The Cast AI 2026 report says that average GPU use on company servers is 5 percent. Almost all of the capacity they buy waits and does nothing. When this empty space is filled with free usage, the company mostly pays only the electricity bill.

And the thing they get in return is more valuable than tokens. OpenRouter’s stealth model program agreement does not even hide this. Some providers offer their models free for a period of time and hide the name. The goal is to collect data, to train and improve the model. The usage is free. The thing they buy is data. The shape of the traffic is not a coincidence either. OpenRouter’s data shows that for every 31 units of input, the model produces only 1 unit of output. The model reads very big code files and writes small fixes. The logic is simple: processing input is much cheaper than producing output.

The model hunt

When a model has no name, the detective game starts at once. Everybody looks for the maker. Some of the methods are really clever.

Error message. All of the models of the Chinese company Zhipu on OpenRouter give one certain error in the same way. Ox Alpha does the same. The critical detail is this: when the same model runs on another platform, it gives a completely different error. So this signature does not belong to the model. It belongs to the platform of the company that runs it.

Leak. A request that was broken on purpose returned a file path in the error output. That path was exactly the same as the path in Zhipu’s own documents.

Video encoder. In four different test videos, Ox Alpha spends almost the same number of tokens as Zhipu’s vision model. None of the other candidates in the test matches this pattern.

Working style. When it fixes a bug, Ox Alpha takes 117 steps on average. Zhipu’s known model works in 124 steps. GPT-5.6 Sol takes 61 steps, and Claude about 90.

Davis says he is 99 percent sure. In prediction markets the Zhipu option has a price of 84 percent. But of course it is not certain. Zhipu publishes its own models with an open license, so anybody can download them and change them. Another company may have taken one of them. In that case most of the signatures above still work. In fact, Cursor’s Composer 2.5 model was built on top of a model from Kimi.

There is also geography. A developer sent the same request from seven different points in the world and measured the response times. Atlanta 57 milliseconds, Dallas 76, Europe about 130, Asia more than 200. The measurements point to the East Coast of the United States. This is not the picture we expect for a Chinese company.

In short, there are two scenarios, and we have no proof that lets us say “we caught it.”

The numbers change

Davis’s test had ten bugs. The full test has 113 bugs, 91 projects and five programming languages. When the whole test runs, the result is about 63 percent. This is much more reasonable. A score of 80 percent means the model is far above all frontier models today. In another test, where the model gets no tools, the result falls to 28 percent. So every test produces results at very different ends.

The same model, different numbers.

Where is the data

There is a point here, and we must be careful about it. OpenRouter’s page says this: the prompts you send and the answers of the model are stored at the provider, and they are not used for training. But OpenCode’s documents say that the provider stores nothing.

Two documents that say the opposite of each other. It seems that the code you send stays at the provider. So you must be careful.

Thursday

The free access will end on 27 August. And we will learn the owner of the model. This is the fifth anonymous model in six months. Chinese companies owned the four models before it: Zhipu, Xiaomi, Ant Group, Meituan. This one is probably Chinese too.

Normally, when a model comes out, we know who made it, and we discuss how good it is. This time it was the opposite. For three days people looked for the maker. They looked at error messages, they counted the tokens that the video encoder spends, they measured how many steps the model thinks in. The result was not an answer. It was a guess.

People also discussed how good it is for three days. There are different numbers, and there is no agreement about which one is correct.

We will learn the owner of the model soon. But the discussion about these measurement and test methods will not end.

Sources

• OpenRouter, “Ox Alpha model page,” 20 August 2026 — openrouter.ai

• OpenRouter, “Stealth Program End User License Agreement,” July 2026 — openrouter.ai

• Bloomberg, “Mystery AI Model Ox Alpha Draws Developers With Free Access,” 23 August 2026 — bloomberg.com

• Ben Davis (@davis7), DeepSWE test and correction, 21-22 August 2026 — x.com

• Datacurve, “DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks,” 2026 — arxiv.org

• Manifold, “Who is behind Ox Alpha, the mysterious new frontier LLM on OpenRouter?” August 2026 — manifold.markets