Ox Alpha appeared in stealth, no company name, conventional model card or price tag. Developers have already sent it *trillions* of tokens.

UPDATE:

(the news is going to come fast and furious on this one) *possible* provenance, while other commentators are making vague comments like “it’s not who you think it is” and a second stealth model has appeared, nicknamed korrine.

A new claim identifies Ox Alpha as the forthcoming GLM-5.3 Flash. The broader attribution to Z.ai’s GLM family is supported by technical fingerprinting, but the specific model designation remains unconfirmed.

source: Twitter/X.

Original Post:

Is it GLM / Z.ai? The AI chattersphere is alive with theories on who is behind the launch of a new powerful anonymous model with impressive numbers and already skyrocketing adoption. An unbranded, unusually capable AI model appeared on OpenRouter on August 20 with almost no explanation of where it came from. Its name is Ox Alpha, although “the Mystery Model” is what we’re going with for now.

It is free, multimodal and built for coding and sustained agentic work. It accepts text, images and video, offers a 1,048,576-token context window and can generate responses as long as 131,072 tokens. Without a publicly identified developer. OpenRouter’s listing says Ox Alpha is operated by an anonymous third-party provider. OpenRouter is only routing requests to the model and says it is not the developer, owner or provider.

The Mystery Model has instantly become something unusual, even in a three million model universe: a frontier-scale product launch without a company, brand or official announcement attached to it. And because it’s numbers are so impressive everyone wants to know where it’s come from.

What we know, and what we do not

The public claims surrounding Ox Alpha are moving faster than the evidence. The anonymity has not discouraged developers. OpenCode’s public usage page reported approximately 2.4 trillion tokens, 66,000 users and more than one million completed sessions shortly after the model appeared.

Provenance: unproven. The leading theory is that Ox Alpha comes from Z.ai, the Chinese AI company also known as Zhipu AI, and represents an unreleased multimodal member of the GLM-5 family. There is decently sufficient evidence behind that theory.

The fingerprints point toward GLM

Independent developer testing has found that Ox Alpha processes video in a remarkably similar way to Z.ai’s GLM-5V-Turbo. Across four controlled video samples, the two models reportedly consumed exactly the same number of tokens. The matching results held as frame rates, duration and resolution changed. Other candidates, including Xiaomi’s MiMo v2.5 and Alibaba’s Qwen models, produced substantially different token patterns.

The same investigation found that Ox Alpha’s token counts matched GLM-5.3 across 25 varied text prompts, apart from a constant 75-token difference that could come from a hidden system wrapper.

Ox Alpha also rejects audio in a manner consistent with GLM-5V, uses a similar output style and appears to take a comparable number of steps when completing long coding tasks.

The full independent fingerprinting analysis assigns approximately 90 percent confidence to the Z.ai theory. Strong evidence, if not proof.

Matching tokenizers and video-processing behaviour could indicate a shared architecture, a related checkpoint or the reuse of common model components. API wrappers can also affect observed token counts. Only the operator or model developer can conclusively claim ownership.

There is, however, precedent. Another anonymous OpenRouter model called Pony Alpha was later revealed as an early version of Z.ai’s GLM-5. Hunter Alpha, initially mistaken for a possible DeepSeek release, was eventually identified as an internal version of Xiaomi’s MiMo-V2-Pro. Reuters described the episode as part of a growing practice of stealth-testing models with real developers before formally revealing their origins. Ox Alpha appears to be the latest and potentially largest version of that experiment.

The timing makes the Z.ai theory more credible

Z.ai released GLM-5.3 on August 14, only six days before Ox Alpha appeared. According to Z.ai’s documentation, GLM-5.3 uses the same base model as GLM-5.2. Its improvements came entirely from post-training, including reinforcement learning conducted in more realistic, long-horizon software environments.

Z.ai claims those changes produced a 50 percent improvement on its internal coding benchmark. GLM-5.3 also has a one-million-token context window and a maximum output length of approximately 128,000 tokens. The public version of GLM-5.3 is text-only. Ox Alpha adds images and video while retaining a suspiciously similar tokenizer and context profile.

There is one more piece of supporting evidence. In June, Reuters reported that Z.ai expected to release GLM-5.5 in August. While this does not make Ox Alpha GLM-5.5. It does make the possibility harder to disprove.

The model could be GLM-5.3V, an advanced GLM-5 checkpoint, a preview of GLM-5.5 or something else built from the same family of components. For now, “an unreleased multimodal GLM” is the strongest hypothesis.

Is it really destroying the benchmarks?

Developer reports describe Ox Alpha as one-shotting coding tasks that other leading models fail. One independent test ran the model against a deterministic ten-task subset of DeepSWE and recorded eight successful completions. While I personally find leaderboard numbers not particularly meaningful they are the numbers the industry lives and dies by.

While impressive, it is not the same as winning DeepSWE. The complete DeepSWE benchmark contains 113 original, long-horizon engineering tasks across 91 repositories and five programming languages. Ox Alpha does not yet appear on its full public leaderboard.

A score of eight out of ten can be an important early signal, particularly when the model solves a task that several competitors repeatedly failed. It remains too small a sample to support claims that Ox Alpha has surpassed Claude, GPT or the existing GLM models.

What does 100 trillion tokens per day mean?

OpenCode says the provider has capacity to serve 100 trillion tokens per day. That would equal roughly 1.16 billion tokens every second. It is an extraordinary number, but the claim lacks some definition. It could include input tokens, cached tokens and theoretical peak capacity. It should not necessarily be interpreted as 100 trillion newly generated output tokens.

OpenCode reports that approximately 94 percent of Ox Alpha’s observed input tokens have been served from cache. That would dramatically alter the infrastructure required and the economics of the headline number.

The fact that an unidentified operator is willing to subsidize an enormous public evaluation, apparently at no cost to users, is fascinating. Free inference is not charity. The provider receives something potentially more valuable: millions of real coding sessions, difficult agent trajectories, failure cases and unbiased comparisons against established models.

And then there’s the how. Ox Alpha is not merely being released. It is experiencing a novel, lower risk way of being pressure-tested by the market, with maximum intrigue.

The arrival of the blind model audition

Traditional model launches are built around brand recognition, curated benchmark charts and carefully managed demonstrations.

Stealth releases reverse that process. Developers encounter the product before they encounter the company. They evaluate whether it can repair a repository, control tools or complete a long-running task without being influenced by the name attached to it. For model developers, this creates a global blind audition.

A laboratory can test infrastructure, observe agent behaviour, identify weak points and measure demand before making a formal announcement. It can also create a wave of speculation that would be difficult to purchase through conventional marketing.

The practice is especially attractive to Chinese AI companies attempting to demonstrate that their systems can compete directly with the leading American models.If developers adopt an anonymous Chinese model because it works, the performance result (and users) arrive before the geopolitical or brand narrative.

There is also a serious enterprise warning

The Mystery Model’s anonymity and policy has risks that should not be ignored. OpenCode describes the provider as following a zero-retention policy. OpenRouter’s model page, however, says prompts and completions are retained by the provider, although they are not used for training. Those statements are not identical.

Until that discrepancy is resolved and the operator is identified, companies will likely not adopt and should be careful about sending Ox Alpha proprietary source code, credentials, customer records, trade secrets or unpublished intellectual property.

An anonymous preview can be useful for experiments involving public repositories and synthetic tasks. It is not yet a substitute for a vendor that can provide contractual data controls, security documentation, service guarantees and legal accountability.

The model may be a mystery. The market signal is not.

If Ox Alpha is revealed as GLM-5.5 or another Z.ai model, it would demonstrate how quickly a laboratory can turn post-training advances into a powerful multimodal agent.

If it belongs to another company, while increasingly unlikely, the conclusion may be even more significant. It would mean the number of organizations capable of deploying frontier-quality agentic systems is larger than the market realizes.

Either outcome supports the same broader thesis: advanced intelligence is becoming easier to access, harder to attribute and increasingly interchangeable at the application layer.

As models become commodities, provenance becomes more valuable. The scarcity is increasingly not being able to access to an intelligent model. Knowing who built it, what happened to your data and whether the operator will still be there once the free preview ends., is the real issue. Because of this corporate option is likely to be very limited until we know more.

Ox Alpha’s provenance is coming, and probably soon. An important realization has already: there’s a new way of launching a model, and one can now generate global adoption (with sufficiently impressive numbers) before anyone knows its name, or who is behind it.