Signals Inbox·August 23, 2026·Frontier AI

Is Ox Alpha the next Gemini?

Ox Alpha is probably not the next Gemini. The rumor was plausible for a few days, but the model’s tokenizer, reasoning controls, multimodal behavior and even backend errors now point much more strongly toward Z.ai’s GLM family.

Q1What is Ox Alpha, and why does everyone think it might be Gemini?

Ox Alpha is currently the AI industry's favorite guessing game: an anonymous frontier model that looks good enough, large enough and strange enough for people to wonder whether Google is hiding the next Gemini in plain sight.

OpenRouter describes Ox Alpha as a model built for coding, long-running agents and production workloads. It accepts text, images and video, has a 1,048,576-token context window, can generate as many as 131,072 tokens in one response, and is free during its stealth preview. OpenCode has also offered near-unlimited access while relaying a provider capacity of 100 trillion tokens per day.

The model quickly attracted serious attention. Stripe CEO Patrick Collison called it “very impressive,” developers started posting strong coding runs, and Google employees began vagueposting while nobody had claimed the model publicly. Google DeepMind currently has an unreleased Gemini 3.5 Pro marked “coming soon,” so the obvious theory wrote itself: perhaps Ox Alpha is Gemini 3.5 Pro being tested anonymously.

There is one big problem with that theory. Since the first wave of speculation, researchers have started probing the model rather than judging how it feels. The picture is much less ambiguous now.

Q2When people call Ox Alpha “the next Gemini,” what are they actually claiming?

The interesting Ox Alpha theory is that Google itself built the model, most plausibly as a stealth preview of Gemini 3.5 Pro.

Simply saying that Ox Alpha is “Gemini-like” tells us very little. Frontier models increasingly converge on the same capabilities: long context, multimodal input, tool use, reasoning and coding. A Chinese model trained partly on Gemini outputs could also behave surprisingly like Gemini without being a Google model.

The actual rumor is much stronger. Google DeepMind still labels Gemini 3.5 Pro as “coming soon,” while Ox Alpha appeared anonymously at almost exactly the point where people were waiting for that missing Pro model. Ox Alpha is also aimed at complex coding and agentic work, which would make sense for a new Gemini flagship.

So this article is testing one specific idea: did Google quietly put an unreleased Gemini on OpenRouter under the name Ox Alpha?

Q3Has Google secretly tested a Gemini model like this before?

Yes. Google has already launched a major Gemini model under a fake name before revealing what it was.

The clearest example is Nano Banana. Google submitted what later became Gemini 2.5 Flash Image to LMArena under the temporary codename nano-banana during pre-release testing. The model went viral before Google revealed its identity, generated millions of Arena votes and eventually became one of Google's most recognizable AI brands.

That precedent gives the Ox Alpha rumor real credibility. An anonymous model becoming unexpectedly popular while Google employees feed the speculation is entirely compatible with Google's previous behavior.

There is also a major difference. Nano Banana was tested anonymously on LMArena, where blind model testing is the product. Ox Alpha is being served directly through OpenRouter as a full API model, with unusually generous free inference and an outside provider retaining prompts and completions. That would be a much more elaborate stealth release for Google.

Google has shown that it enjoys this kind of launch. We still need evidence that this particular model is Google's.

Q4Are the cryptic Google tweets actually pointing to Ox Alpha?

The Google posts are the best evidence for the Gemini theory, but they currently look more like deliberate teasing than a genuine model leak.

The post that really accelerated the rumor came from Evan Otero, who works around Google AI Studio and the Gemini API. He wrote: “What if the Ox Alpha was the friends we made along the way.” Around the same time, other Google-linked accounts were posting cryptic lines such as “Google is back. Trust the process.”

Convenient timing, obviously. Google also knows exactly what happened with Nano Banana, so employees participating in another mystery-model frenzy naturally gets attention.

But the latest developments make those posts harder to interpret as breadcrumbs. AI researcher Mia, who had previously claimed to know Ox Alpha's identity, later responded directly to Otero's post by saying Ox Alpha was not a new Gemini. That statement is itself unverified, but it shows how quickly the social-media clues can point in opposite directions.

The Google posts deserve some weight. Not much compared with evidence coming from the model itself.

Q5Do Ox Alpha's specs actually look like Gemini?

Ox Alpha's headline specs could belong to Gemini, but the weirdly specific details line up much better with Z.ai's GLM family.

Google's current Gemini 3.7 Flash supports a 1,048,576-token input window, so Ox Alpha's one-million-token context proves very little. Both are multimodal and both are designed for agentic work.

Further down the spec sheet, it gets more interesting. Gemini 3.7 Flash currently stops at 65,536 output tokens and exposes low, medium and high thinking levels. Ox Alpha allows 131,072 output tokens, always reasons, and uses low, high and max.

OpenRouter's current GLM-5.3 listing has the same 1,048,576-token context as Ox Alpha, the same 131,072-token output ceiling, mandatory reasoning, the same low/high/max scale and the same max default. Meanwhile, Z.ai's GLM-5V Turbo already accepts the same three input types as Ox Alpha: text, images and video.

We do not know the specs of unreleased Gemini 3.5 Pro, so none of this rules Google out by itself. It does show that Ox Alpha happens to combine several very specific properties already found across Z.ai's latest models.

Q6Does Ox Alpha's tokenizer give away where it came from?

The tokenizer is currently the cleanest technical clue we have, and the latest testing ties Ox Alpha extremely closely to GLM-5.

A tokenizer is the machinery that breaks text into pieces before a model processes it. Different model families split unusual words, punctuation, Unicode, code and different languages in slightly different ways. That makes tokenization surprisingly useful for fingerprinting an anonymous model.

The earliest Ox Alpha experiments used only a handful of prompts and noticed that its token counts tracked GLM models unusually closely. That was interesting but still easy to dismiss.

The larger test is much harder to wave away. Researcher Joseph W. Elstner tested 95 deliberately chosen strings against 14 candidate vocabularies. An older GLM vocabulary matched 84 of 95. The best non-GLM candidate managed 46.

Once the released GLM-5 vocabulary was added, Ox Alpha matched 95 out of 95 probes, with zero average error.

A second 29-input test compared Ox Alpha with Z.ai's served GLM endpoint and a local tokenizer implementation. They agreed on 25 inputs; the four differences involved system-only marker strings where the hosted services applied additional input handling.

That still does not tell us who owns Ox Alpha. Anyone can theoretically use a public tokenizer. But the idea that Ox Alpha simply happens to resemble GLM has become very difficult to defend. Its text is being chopped up exactly the way GLM-5 chops it up.

Q7Why are Ox Alpha's reasoning controls such a useful clue?

Ox Alpha also copies several unusually specific GLM-5.3 API choices, which makes the tokenizer match harder to treat as an isolated coincidence.

OpenRouter currently says GLM-5.3 cannot turn reasoning off. It accepts low, high and max, with max as the default.

Ox Alpha behaves the same way. Attempts to disable reasoning have produced an error telling users that the model always thinks and that they should choose low, high or max.

Google's current public interface is different. Gemini 3.7 Flash supports low, medium and high. Its normal ladder therefore contains medium and stops at high, while Ox Alpha skips medium and adds max.

There is no engineering law forcing two labs to make those same arbitrary choices. When those controls appear alongside the GLM tokenizer and the matching 131,072-token output limit, a GLM connection becomes a much simpler explanation.

Q8What do the image and video tests say about Ox Alpha?

The visual side of Ox Alpha also appears to behave like Z.ai's multimodal models, which closes one of the biggest gaps in the GLM theory.

Public GLM-5.3 is text-only, while Ox Alpha accepts images and video. That initially looked like evidence against GLM.

Researchers then started measuring how Ox Alpha converts visual inputs into tokens. Controlled image tests across different resolutions reportedly followed the same patch geometry associated with GLM's vision stack. Separate video experiments changed duration, resolution and frame rate one variable at a time; Ox Alpha and GLM-5V Turbo reportedly consumed matching token budgets across those tests.

That gives us a plausible explanation for why Ox Alpha is more capable than public GLM-5.3 without requiring a completely different model family. Z.ai already has native text-image-video architecture in GLM-5V Turbo. Its newest reasoning model, GLM-5.3, currently lacks that multimodality.

An unreleased model combining the two would look a lot like Ox Alpha.

Q9Did Ox Alpha accidentally expose Z.ai's own backend?

The backend evidence currently makes a direct Z.ai connection much more plausible than a shared tokenizer alone would.

One researcher deliberately sent malformed input through Ox Alpha and received a Java stack trace containing com.wd.paas.api.domain.v4.chat.ChatCompletionRequest. That path closely mirrors Zhipu's documented /api/paas/v4/chat/completions structure.

Another malformed request generated a code 1214 error envelope matching the format returned by Z.ai-hosted GLM models.

The useful control came from running GLM weights through another host. That host returned a different validation format. In other words, the matching error behavior seems to come from the server stack rather than from GLM's model weights.

That distinction matters here. A Microsoft or Google model could theoretically reuse GLM's public vocabulary. Accidentally surfacing Z.ai-style internal API plumbing at the same time requires another explanation.

We still cannot call that official attribution. Z.ai has not claimed Ox Alpha, and a partner could be serving a derivative through Z.ai infrastructure. Yet today this is probably the hardest evidence for the Gemini theory to explain away.

Q10Could all these GLM clues still be misleading?

Yes, although a third party building on GLM now looks much more plausible than Google secretly building Ox Alpha from scratch.

The tokenizer tells us model lineage rather than corporate ownership. The API behavior tells us something about implementation. The backend errors tell us something about whoever is serving it. Those three questions do not necessarily have the same answer.

A company could take GLM technology, post-train it substantially, add multimodality, and serve it through Z.ai infrastructure. An American company could therefore own or finance the finished product while Ox Alpha still leaves GLM fingerprints everywhere.

So we would not say that Ox Alpha has been “proven to be GLM-5.3.” Public GLM-5.3 only accepts text, whereas Ox Alpha accepts images and video. Ox Alpha may be another GLM checkpoint, a newer multimodal variant, or a product built on top of GLM technology.

The evidence currently narrows the family much more confidently than it identifies the exact model name.

Q11Could Ox Alpha actually be Microsoft's next MAI model?

Microsoft is the newest serious Ox Alpha theory, but right now it explains the available evidence worse than a GLM-based model does.

The Microsoft idea gained traction after researchers noticed tokenizer behavior that could be connected to OpenAI-style vocabulary ancestry, and Business Insider reported that speculation had begun shifting toward Microsoft's upcoming MAI models. Microsoft also has an obvious advantage over Z.ai when people ask who could afford to give away this much inference.

The theory becomes more interesting if we separate architecture from ownership. Microsoft could theoretically be testing a model built partly on another model family, or an outside company could be operating a GLM derivative on very large Western infrastructure.

However, the newest tokenizer work weakens the simple version of the Microsoft theory. Ox Alpha did not merely look vaguely compatible with an OpenAI-derived tokenizer. It matched the released GLM-5 vocabulary on all 95 probes. The Z.ai-style server errors create another problem that Microsoft's compute budget does not solve.

So Microsoft deserves a place in the probability table now. The evidence for a pure Microsoft MAI model remains thin.

Q12Is Ox Alpha actually better than the frontier models we already have?

Ox Alpha currently looks like a strong frontier model, but the first viral benchmark made the gap look much larger than subsequent testing suggests.

The number that drove much of the early excitement was 80% on DeepSWE. Ox Alpha solved eight of ten software-engineering tasks in one early run, ahead of the comparison models used in that test.

Ten tasks are nowhere near enough for an 80% result to carry much weight. One additional failure moves the score by ten percentage points.

Larger runs quickly changed the picture. One full 113-task DeepSWE run reported 58.4%, roughly alongside Claude Opus 4.8 at 59%. Another full community run landed around 63%. Different harnesses and settings can explain some of that spread, but both results tell essentially the same story: Ox Alpha looks competitive with frontier coding models rather than obviously one generation ahead of them.

Other tests have also been mixed. Developers have posted excellent agentic coding examples, while a 175-problem LiveCodeBench run reported 28% Pass@1 and private evaluations have found weaker results in some tasks. Several developers also say the harness matters a lot, with Ox Alpha behaving better inside Hermes than OpenCode.

The model is clearly interesting enough to investigate. Its performance no longer gives us any reason to assume that only Google could have built it.

Q13Why would Z.ai hide another model behind a strange name?

Z.ai has already done almost exactly this once, which makes an Ox Alpha stealth test feel quite normal.

Earlier this year, an anonymous model called Pony Alpha appeared on OpenRouter. It was strong at coding and agentic work, developers tried to reverse-engineer its identity, and Z.ai later confirmed that Pony Alpha had been an early testing version of GLM-5. OpenRouter's old Pony Alpha page now explicitly says the same thing.

That history is hard to ignore. Ox Alpha appeared on the same platform, uses another animal-plus-Alpha name, targets coding and long-running agents, and has now accumulated technical fingerprints pointing back toward the same model family.

If Ox Alpha does belong to Z.ai, the most obvious product explanation would be a new multimodal GLM-5.x model. Public GLM-5.3 already provides the million-token context and newer reasoning stack. GLM-5V Turbo already handles text, images and video. Ox Alpha currently looks like the point where those two branches meet.

We cannot know whether Z.ai would call that GLM-5.3V, GLM-5.5 or something else. Guessing the suffix adds false precision. The broader GLM-5.x hypothesis is much stronger.

Q14What would make us believe Ox Alpha really is Gemini?

We would take the Gemini theory seriously again if Ox Alpha started producing Gemini-specific technical fingerprints rather than Google-themed social-media clues.

A direct Google reveal would obviously settle it. Before that, we would want something such as Gemini-specific tokenization, Google's visual preprocessing behavior, an internal error trace tied to Google's serving infrastructure, or another architectural property that GLM models cannot explain.

A much more coordinated teaser campaign from the Gemini team would also move the probability. With Nano Banana, the vague clues eventually became part of a recognizable reveal campaign around an anonymous model that Google genuinely owned.

More impressive Ox Alpha demos would not change much. Neither would Ox Alpha calling itself Gemini, because its hidden instructions explicitly tell it to identify itself only as ox-alpha from an undisclosed organization.

Right now the Gemini case is missing the one thing the GLM case increasingly has: fingerprints from inside the system.

Q15So who most likely built Ox Alpha?

Our current best estimate is an 80% chance that Ox Alpha is an unreleased Z.ai GLM model, with another 14% reserved for somebody else building or operating a heavily GLM-derived model.

These percentages are our judgment from the available evidence rather than measured probabilities. We are deliberately separating “GLM technology” from “Z.ai owns the final product,” because the technical tests support the first claim even more strongly than the second.

The Microsoft theory deserves a few percentage points because it is recent, Microsoft can clearly support the required inference scale, and hybrid ownership or infrastructure arrangements are possible. Gemini has fallen to roughly 2% for us. Google has given people excellent reasons to speculate but, so far, almost none of the technical clues have moved in its direction.

As seen above, the same GLM connection has now appeared independently in the tokenizer, API behavior, multimodal processing and serving stack. At this point Google would need a surprisingly complicated explanation for all four.

Q16Is Ox Alpha actually the next Gemini?

Probably not. Based on everything we can test today, we would put the chance that Ox Alpha is a secret Gemini at roughly 2%.

The rumor made much more sense before people started taking the model apart. Google has a real history of anonymous pre-release testing with Nano Banana. Gemini 3.5 Pro is currently still waiting in the wings. Google employees then joined the Ox Alpha conversation at exactly the moment speculation was taking off.

But the evidence has moved quickly since then. Researchers now have a 95-of-95 match between Ox Alpha and the released GLM-5 tokenizer. Its reasoning controls closely follow GLM-5.3. Visual tests point toward the GLM multimodal stack. Malformed requests have even exposed server behavior associated with Z.ai infrastructure.

None of those findings gives us the name printed on the future model card. Together, they make Gemini a very unlikely answer.

There is a more interesting possibility here anyway. If Ox Alpha eventually turns out to be another Z.ai model, thousands of developers spent several days wondering whether a Chinese model was secretly Google's next Gemini.

That would say quite a lot about how narrow the frontier has become.