By ByteBot
A model called Ox Alpha appeared on OpenRouter four days ago. It is free, handles one million tokens of context, reportedly beats GPT-5.6 on a coding benchmark, and nobody knows who built it. In three days it reached 221,000 users and 16 trillion tokens of usage. Most of those developers pasted real code into it without asking where it goes.
What Ox Alpha Is
Ox Alpha is a “stealth model” under OpenRouter’s preview program — a format where anonymous third-party providers release models free for roughly a week before an official launch. The model appeared August 20 and the free window closes around August 27. After that it either gets a named, priced release or vanishes.
The specs are legitimately impressive: a 1,048,576-token context window matching Claude Fable 5, multimodal input (text, image, video), function calling, 22 tokens per second throughput, and 99.98% uptime in its first three days. The price is zero — which explains the explosion in adoption.
Why Developers Are Excited About the Ox Alpha Model
Developer Ben Davis ran a 10-task subset of the DeepSWE benchmark and clocked Ox Alpha at 80% Pass@1, against 65% for Claude and 52% for GPT-5.6. That number went viral. However, the caveat matters: it was a 10-task subset, not a full audit. Community testing on the full suite produced roughly 63%, more in line with GPT-5.6 Sol mid. Ox Alpha does not yet appear on the official DeepSWE leaderboard, where Claude Opus 5 leads at 73.6%.
The caveats matter, but they do not erase the signal. Nous Research’s Hermes Agent integrated Ox Alpha immediately. So did the Zed code editor. Stripe CEO Patrick Collison called it “very impressive.” Moreover, in agentic testing it made a single error across 69 consecutive tool calls. For developers building agent loops, that coherence at depth is the actual value.
Who Built It — And Why That Matters
OpenRouter does not disclose the provider. That is by design: the Stealth Program explicitly guarantees provider anonymity. Nevertheless, community researchers have done forensic fingerprinting and arrived at a 98% confidence attribution to Zhipu AI’s GLM-5.3 family. The evidence: the tokenizer matches GLM-5.3 exactly with only a fixed 75-token wrapper offset; the video encoder token consumption rate matches GLM-5V-Turbo precisely; the output emoji frequency aligns with GLM and Qwen models, not with Claude, GPT, or Grok; and an API stack trace names a path consistent with Zhipu’s documented infrastructure.
None of this is confirmed. No lab has claimed the model. The pattern, however, is familiar. Chinese AI labs — Zhipu, Xiaomi, Ant Group, Meituan — have released stealth models on OpenRouter before official announcements multiple times in 2025 and 2026. The free week generates eval traffic, benchmark data, and developer mind-share before the paid product launches. TechCrunch noted that the mystery has driven significant speculation across developer communities.
The Part Most Developers Skipped
OpenRouter’s Stealth Program End User License Agreement grants both OpenRouter and the anonymous stealth provider the right to collect, retain, and process user content. The Ox Alpha model page adds: “Prompts and completions are retained by the provider and are not used for training.” That is a policy claim from an entity you cannot identify, audit, or issue a data erasure request to.
OpenCode, which also hosts Ox Alpha, advertises “Zero Data Retention” at its routing layer. That does not bind the upstream anonymous provider. The 221,000 developers who ran 16 trillion tokens through this model in three days sent their code somewhere. That somewhere does not have a name attached to it.
For European developers this is not just a privacy preference — it is a compliance problem. The EU AI Act, in force since August 2, requires provider identification and output watermarking. GDPR mandates a named data controller before any data processing agreement is possible. An anonymous provider fails both requirements. Furthermore, until Ox Alpha’s creator identifies itself, European enterprises cannot legally route production data through it.
What to Do With the Ox Alpha Model
The model is worth testing. It is genuinely capable, the 1M context window is useful for large codebase tasks, and free compute at this quality level does not appear often. The rule is straightforward: treat it like a public endpoint with unknown data practices, because that is exactly what it is.
- Use for: synthetic code generation, open-source work, public-domain tasks, anything you would post on GitHub
- Avoid: proprietary source code, customer data, API keys, internal architecture documents, anything confidential
- Timeline: free window closes around August 27 — if the provider reveals itself and publishes auditable data handling terms, reassess then
The Hacker News community’s top-voted reaction said it plainly: “who the fuck made this model and where did they get this much compute?” That question is not rhetorical. It is the one worth answering before you run your next agent loop through it.
