An unnamed model has the entire AI developer community playing a guessing game.

An anonymous AI model called Ox Alpha quietly appeared on the OpenRouter platform, free to use and remarkably powerful, quickly sparking heated discussions among developers. Chinese netizens, noting that "Ox" translates to "牛" (ox/cow), affectionately nicknamed it the "Niu Lai" (Here Comes the Ox) large model.

Free and powerful, developers flock to it

Late on August 20th US Eastern Time (early morning August 21st Beijing time), Ox Alpha was listed on OpenRouter in stealth mode, provided by an anonymous third party.

Its specifications are impressive: a 1 million token context window, support for text, image, and video multimodal inputs, full tool calling support, and forced reasoning. The price is even more astonishing—completely free.

According to reports, Ox Alpha is free for one week with "near-unlimited" usage. The AI coding tool OpenCode officially announced that its provider has access to 100 trillion (100T) tokens of computing power per day—roughly 100 times the entire monthly AI token usage of a company like Visa.

Stripe CEO Patrick Collison tested it personally and wrote on X: "Very impressive."

Coding tests surpass GPT-5.6 and Claude

davis7, the author of the AI coding framework fable, ran 10 problems using the official DeepSWE evaluation and was stunned by the results. He wrote on X: "I am very confused."

The data speaks for itself: Ox Alpha scored 80%, Claude Fable 65%, and GPT-5.6 Sol 52%.

An anonymous model beat two well-known closed-source frontier models by 15 percentage points. The community later independently retested, with results largely consistent: 4 passed, 4 near-passed, and 2 failed, aligning with the 8/10 outcome.

Two independent tests point to the same conclusion: this mysterious model's coding ability is indeed at the frontier level.

A 24-hour fingerprint detective battle: five clues point to Z.AI

With its identity unknown, the community began digging.

According to a detailed analysis on X by lizikk_zhu, founder of Atomstorm.ai and former executive at Baidu and Alibaba, the community took just 24 hours to identify the prime suspect through technical fingerprint comparison of public APIs.

The strongest clue came from the video encoder. Feeding the same controlled test videos to different models and recording their token consumption, Ox Alpha and Z.AI's GLM-5V-Turbo matched exactly token-for-token: 296, 296, 884, 1064, without a single difference. The analysis noted this matched three independent design features—frame sampling strategy, duration scaling, and resolution scaling—making the probability of coincidence "virtually zero."

Four additional clues emerged:

Text tokenizer: 25 special strings (including CJK, emoji, and Cyrillic characters) with token counts identical to GLM-5.3.

Error codes: passing incorrect parameters returns error [1210]—an error code analysts say is unique to Z.AI.

Chinese backend: when triggering image parsing errors, the backend directly returns Chinese text: "图片输入格式/解析错误" (Image input format/parsing error).

Content moderation: answers to sensitive questions follow an official script.

These five clues come from completely different technical layers, yet all point to Z.AI. lizikk_zhu concludes that Ox Alpha is highly likely to be Z.AI's unreleased GLM-5.3 vision version (or GLM-5.5), with 70-80% confidence.

Tech media have also noted that Z.AI previously tested GLM-5 anonymously under the codename "Pony Alpha," and developers have found similar tokenizer behavior and response patterns in Ox Alpha.

But the answer may not be Z.AI

Doubts remain.

First, the computing power doesn't add up. With 100T tokens per day free for a week, multiple analyses suggest that Z.AI's current computing capacity could not independently support this. The community has thus proposed a "synthesis hypothesis": GLM weights plus big-tech computing power hosting—the model may be Z.AI's, but the servers could belong to Xiaomi or Tencent.

Second, someone has hinted at a twist. Mia, a well-known AI industry leaker, posted on X: "I've confirmed who Ox Alpha belongs to, but I can't say yet. You'll all be very excited—it's not who you think." A flood of question marks followed.

AI analyst Andrew Curran wrote on X that on Friday night GLM was the mainstream guess, but by Saturday morning, "everyone seems less sure about anything."

Another analysis suggests that Ox Alpha's tokenizer features could also point to Microsoft's MAI series models.

The mystery will likely be resolved on August 28th—when Z.AI officially open-sources the GLM-5.3 weights, and the model card's release may reveal the identity once and for all.

Where does the data go? Contradictory retention policies

Beyond the identity guessing game, a more practical question remains unresolved: the fate of user data.

According to lizikk_zhu's analysis, the OpenRouter model card explicitly states: "Prompts and completions are retained by the provider, not used for training."

Meanwhile, OpenCode's official announcement claims: "Zero Data Retention."

Two official channels, two different statements—at least one is incomplete. "Not used for training" and "zero data retention" are two distinct concepts: the former means conversations remain with the provider with opaque retention periods; the latter means deletion immediately after request completion, auditable in the corporate compliance sense.

Anonymous launches become an industry norm

This is not an isolated phenomenon. Companies like ByteDance, Sea, and Alibaba have all released models this year without disclosing their source, allowing them to test new services with fewer constraints.

The timeline shows this playbook has multiple precedents: Hunter Alpha was later confirmed to be Xiaomi's MiMo-V2-Pro, Owl Alpha was Meituan's LongCat 2.0, Pony Alpha was Z.AI's GLM-5, and even OpenAI tested GPT-4.1 under the codename Quasar Alpha.

The logic is clear: anonymous launches mitigate brand risk, free access attracts real production traffic, stress-test feedback is more valuable than internal testing data, and finally, claiming the model upon release completes the monetization cycle.

If Ox Alpha ultimately scores 61 on the AA Intelligence Index as analysts predict, it would become the first open-weight model to break the 60-point barrier, directly touching the threshold of closed-source frontier models.