Ox Alpha: The Free Mystery AI Model Disrupting AI Coding

No name, no company, no invoice

On Thursday, August 20, 2026, a new model appeared on OpenRouter under the identifierstealth/ox-alpha. No developer attribution, no research paper, no launch tweet. Price: zero. Within three days it had drawn more than 220,000 users and served over 220,000 users, according to independent coverage of the platform. (Earlier drafts of this story repeated a "16 trillion tokens in three days" figure from secondary coverage; we have removed it because no primary source confirms it — verified reporting puts usage at roughly 2 trillion tokens in its first 48 hours.)

The reason is simple: the specs on its listing are not those of a throwaway experiment. OpenRouter describes Ox Alpha as "a reasoning model designed for coding, sustained agentic work, and production workloads" — aimed squarely at long-horizon software engineering, not casual chat.

Confirmed facts: what we actually know

These points come straight from OpenRouter's official model page — the only source you can treat as verified:

• A 1,048,576-token (1M) context window with roughly a 131K-token output limit.
• Text, image, and video input — the first anonymous release to advertise video input; output is text-only.
• Completely free during the preview: $0 per million tokens in and out.
• An anonymous developer: OpenRouter states explicitly that it is a third party that chose to stay unnamed during the preview, and that OpenRouter is not its developer or owner.
• Also available through OpenCode Zen, which offered it free for one week at near-unlimited usage, under the identifier x-preview-f-free.

Who is behind Ox Alpha? The GLM fingerprint

With the developer hidden, the community turned detective. Developers compared token counts against known models, probed error messages, and analyzed writing style. The strongest public findings:

• An adversarial string battery (CJK, emoji, Cyrillic) matched Ox Alpha's token counts to the GLM tokenizer in 13 of 13 cases, and a 25-prompt sweep aligned with GLM-5.3 within a constant wrapper offset.
• A rejection carrying error code [1210] when disabling thinking mode was verbatim identical to Z.ai's GLM-5.3 API error — same number, same wording.
• The open-source modelprint fingerprinting tool matched Ox Alpha to GLM-5.3 on six of nine infrastructure probes; the next-best candidate cleared two.
• An independent gzip compression-distance analysis classified most outputs as closest to GLM-5.3, and one widely shared report quoted the model — after its own system prompt was fed back to it — saying "I'm GLM, made by Z.ai."
• Precedent matters too: Z.ai previously previewed GLM-5 anonymously under the name Pony Alpha before its official reveal.

Important caveat: all of this is strong community evidence, not an official announcement. Even the analysts themselves stress that "matching fingerprints prove shared infrastructure, not identity," and a vocal minority argues for alternatives like Xiaomi's MiMo. Until someone confirms on the record, treat the Z.ai attribution as the leading theory — nothing more.

Does it really beat GPT and Claude?

The viral number — Ox Alpha scoring 80% on a coding benchmark versus 65% for Claude and 52% for GPT-5.6 — came from a run over just ten tasks of the suite. That is far too small a sample for a verdict. Larger independent runs landed closer to 58–63%: genuinely competitive frontier performance, not the blowout social media imagined.

The practical takeaway: the model is clearly strong at long-horizon coding work and worth serious evaluation — but any final ranking againstChatGPT, Claude, orGemini needs broader independent benchmarks, which have not finished landing yet.

The real price: your prompts are the currency

Why would anyone hand out a frontier model for free? Because the accompanying documents say so plainly: free access in exchange for your user content. And here lies the uncomfortable contradiction — the model page says conversations are "retained by the provider and are not used for training," while another OpenRouter document grants stealth providers rights to use user content for training and improvement, with no opt-out. The provider itself is anonymous: no named entity to audit, no one to send a data-erasure request to. Analysts have warned that European companies in particular cannot route production data through it without potentially breaching GDPR and EU AI Act requirements for identified providers.

Why it matters

• The gap is closing fast: an anonymous model entering the "best coding model" conversation means frontier AI is no longer the preserve of big American labs.
• Reverse marketing works: a mysterious free drop achieved adoption in days that marketing budgets struggle to buy — and this playbook is now repeating.
• Data privacy became the core question: when the payment is your data, knowing who hosts a model and what they do with your inputs stops being a nice-to-have. It is now a professional requirement.
• A genuine free trial: a 1M-token window with video input, free of charge, is a rare chance to test frontier capabilities — if you handle your data knowingly.

When does Ox Alpha make sense for you?

• Evaluation and experimentation: probing what frontier models can do on long coding tasks before committing to a paid tool such as Cursor.
• Non-sensitive personal projects: learning code, prototypes, quick ideas worth testing.
• Honest comparison: run it against the tools you already use and judge for yourself — our head-to-head DeepSeek vs ChatGPT review shows how cheap high-performance models differ in practice.

If you handle customer data, proprietary code, or anything that must not leave your control, this is not the right model — use a platform with known provenance and explicit policies instead.

Frequently asked questions about Ox Alpha

What happens next?

The likeliest script, based on precedent: within days to weeks the developer reveals itself — either OpenRouter renames the provider or an official announcement lands — and the model appears under its real identity with real pricing. If it turns out to be an unreleased multimodal GLM build, Z.ai enters the text-plus-video race directly. Watch three signals: the provider name changing on the listing, post-preview pricing appearing, and broad independent benchmarks publishing.

We will update this article the moment anything is confirmed — in AI news, tracking what is verified versus what is hype is the whole job.