An official announcement, but no company
The coding agent platform OpenCode released a new model, Ox Alpha, from its official account on August 21. True to the name "stealth model," the announcement names no developer anywhere. Several outlets reported that it was also listed on OpenRouter the same day, with the provider shown only as stealth. It's an anonymous test deployment, not a formal launch.
On X, developer baba yaga (@babayagatwt) quoted the announcement and wrote that "nobody knows who made it," which fueled the buzz. The replies that followed included impressions from people who had actually tried the model: some said it is "pretty slow," others "usable speed." Both are just individual users' subjective impressions, not official performance metrics.
1 million tokens, and 100 trillion tokens a day
Here are the specs confirmed from the official OpenRouter listing. The context window is 1,048,576 tokens and the maximum output is 131,072 tokens, and it accepts text, image, and video input together. It is presented as a reasoning model aimed at coding, long-running agent tasks, and production workloads, and it states that prompts and outputs submitted during the trial period will not be used for training.
OpenCode said it has "capacity to process 100 trillion tokens a day (100T tokens/day)." That is roughly 1.16 billion tokens per second, a scale that only a handful of places in the world can support. A claim that it "beat Fable, GPT, and Opus on some tests" also appeared in baba yaga's post, but which benchmarks were used was not disclosed. That part has to stay an unconfirmed individual claim.
Why many suspect GLM
Guesses about the maker fall into two camps. baba yaga points out that Xiaomi's in-house MiMo model already has a 1 million-token context window and a multimodal agent, and that the throughput of the "100 trillion token program" Xiaomi ran in collaboration with OpenCode overlaps with this announcement. The other camp raises the possibility that it is the next version of GLM, based on technical circumstantial evidence that the tokenizer and video encoder behave similarly to the Zhipu AI (GLM) family. Both guesses are only circumstantial evidence, and neither company has officially confirmed or denied it.
In the end, only three things are certain right now: a real model has been made free for a week, its specs are officially registered on OpenRouter, and no one has officially said who made it. If you're curious, trying it yourself during the trial period is faster than following the speculation.