Nobody will confirm who built the viral AI model Ox Alpha

A mysterious new AI model called Ox Alpha appeared on OpenRouter on Thursday 21 August, listed as "a reasoning model designed for coding, sustained agentic work, and production workload," according to TechCrunch. It was offered for free, and OpenRouter's own listing described it as a "stealth model," adding that it was "developed and operated by a third-party provider who has chosen to remain anonymous during this preview." The coding platform OpenCode carried the same free release. Officechai reported that the two platforms together made roughly 100 trillion tokens of daily usage available for free during the first week, an unusually generous allowance that has become part of the mystery itself.

The model quickly drew attention from people who matter in the industry. On X, Stripe CEO Patrick Collison — whose company is in the process of acquiring OpenRouter — called Ox Alpha "very impressive," TechCrunch reported. Officechai relayed early user impressions placing the model roughly on par with GPT-5.6 Sol and Fable 5, with one user claiming it beat both on the Deep SWE coding benchmark. None of that, however, answered the one question the release itself refused to: who actually built it.

The tokenizer clue that points to Z.AI

Much of the amateur detective work has focused on the model's tokenizer — the internal system that decides how text is broken into sub-word units before a model processes it, which tends to leave a distinctive fingerprint from lab to lab. Officechai reported that jailbreak researcher Pliny the Liberator analyzed Ox Alpha's tokenizer and concluded it matches the GLM-5.X family built by the Chinese company Z.AI. A more systematic test came from the AI commentator Sushaanth Srinivasan, who ran eleven separate tokenizer probes across ten different models: Ox Alpha matched GLM's behavior on all eleven, while no other lab's models matched on more than four. On one probe involving how a model encodes digits, DeepSeek needed 98 tokens where GLM needed only 29; Ox Alpha needed 29 as well. Sushaanth Srinivasan also found Ox Alpha capable of reading images, which cuts against a pure GLM match, since Z.AI's GLM-5.3 has no vision endpoint at all.

AI analyst Andrew Curran captured how quickly the consensus wobbled, posting on Friday that initial speculation had settled on GLM but that "this morning people seem less sure of anything," TechCrunch reported. Part of the doubt comes from economics rather than linguistics: several commentators argue that the sheer scale of free usage Ox Alpha is absorbing — hundreds of billions of tokens a day, by some estimates — is evidence against a Chinese lab, on the theory that Z.AI is unlikely to have that much spare compute to give away for free.

Microsoft, Google, and two more contested theories

A rival theory names Microsoft. Robert Lukoszko, CEO of the startup Stormy, argued in a widely shared post that a different tokenizer reading changes the picture entirely.

Based on tokenization, stealth/ox-alpha is a cl100k_base-tokenizer model — which rules out OpenAI, Google, Anthropic, xAI and every Chinese frontier lab and matches only Microsoft's Phi/MAI lineage (with IBM Granite as the outside bet) — so it is most likely Microsoft's unreleased MAI frontier model, i.e. 'MAI-2.'

Officechai noted that Microsoft recently shipped a capable image-generation model and is among the few companies with enough spare cloud capacity to fund a giveaway of this size, which lends the theory some financial plausibility even without a matching tokenizer consensus. A third camp points to Google, fresh off releasing Gemini 3.7 Flash to reviews that exceeded expectations while the larger Gemini 3.5 Pro remains delayed; some Google employees reportedly "vagueposted" around the time of Ox Alpha's release, feeding the theory. Officechai reported a complication, though: Ox Alpha is said to have failed a popular vision test that Gemini models typically pass, which cuts against Google as the source. A fourth theory holds that Ox Alpha is Cursor's unreleased Composer 3 model, potentially served through spare compute rented from xAI, which some estimates put at 100 trillion tokens of daily capacity. A fifth names ByteDance: Officechai reported that Ox Alpha served 1.25 trillion tokens on OpenRouter and 2.6 trillion on OpenCode in a single day — 3.85 trillion combined, nearly five times GLM-5.3's entire OpenRouter volume — while staying stable throughout, a scale some argue only ByteDance's Seed infrastructure could sustain, even as the tokenizer evidence keeps pointing back toward the GLM family.

• Z.AI / GLM — tokenizer probes match GLM on 11 of 11 tests, but the free-compute volume looks too large for the theory to some.
• Microsoft — a cl100k_base tokenizer reading points to the Phi/MAI lineage, a possible 'MAI-2.'
• Google — timed near Gemini 3.7 Flash's launch, but Ox Alpha reportedly fails a vision test Gemini passes.
• Cursor Composer — theorized as Composer 3, possibly served via rented xAI compute.
• ByteDance — daily volume (3.85 trillion tokens combined) some argue only ByteDance's Seed compute could sustain.

A conclusion nobody can reach

Reddit users are just as split as the professional commentators: TechCrunch cited one post declaring Ox Alpha "can't be the Chinese," and another expressing "high confidence" that it is. No lab named in any of the five theories has confirmed or denied building Ox Alpha, and OpenRouter's listing still describes the provider only as anonymous. What is not in dispute is the scale of real usage the model is absorbing for sustained agentic work — the same territory OpenRouter's own description points to — while its origin stays unverified.

TechCrunch's reporting closes on a point that outlasts the guessing game itself: whoever built Ox Alpha, the stealth-launch format has generated an outsized amount of attention for a model nobody can name, and other labs may be tempted to copy the playbook rather than announce their next release in the usual way. For developers who have spent the past days routing coding and agentic workloads through Ox Alpha because it is free and fast, that trade-off is worth naming plainly: the model's provider, its data-handling practices, its long-term availability, and even the terms under which it was trained are all currently unknown. A model this capable, deployed into sustained agentic work without a confirmed operator behind it, is a reminder that speed and anonymity can arrive as a package deal — and that the convenience of a free, powerful tool is not the same thing as knowing who is actually running it.