Ox Alpha: Inside the Mystery AI Model That Appeared Out of Nowhere With a 1M Token Context Window

A free, anonymous stealth model just landed on OpenRouter, out-benched several frontier names on early coding tests, and nobody knows for certain who built it. Here's everything worth knowing before the free window closes.

On August 20, 2026, a new model quietly showed up in OpenRouter's catalog. No launch post. No press release. No company name attached — just a provider label that reads "Stealth" and a model ID: stealth/ox-alpha.

Within 48 hours, developers were running it against real coding tasks, tearing apart its tokenizer, and comparing its video-encoding behavior against known models to guess where it came from. That's the world we're in now: frontier-grade models sometimes launch not with a keynote, but with a shrug and a free API key.

Here's a full breakdown of what Ox Alpha actually is, what it can do, why the community thinks it's tied to a specific Chinese lab, and what to make of the benchmark claims currently circulating.

Quick Facts

What Ox Alpha Actually Is

Ox Alpha is being marketed, informally, as a frontier-class reasoning model built for coding, long-running agent workflows, and production use. It's available through three channels right now: OpenRouter, OpenCode Zen, and Mercury Cloud, and it's already been picked up by early adopters like Nous Research's Hermes Agent and the Zed code editor as an available model option.

The headline spec is the context window: 1,048,576 tokens, which puts it in the same bracket as Claude Opus 5, GPT-5.6 Sol, DeepSeek V4 Pro, and Gemini 3.7 Flash — all of which also sit around the one-million-token mark in 2026. To put that in perspective, a million tokens is roughly enough to hand the model an entire mid-sized codebase, or several full-length novels, in a single request.

It also accepts video as input, which is not something most coding-focused models offer. Combined with tool calling and structured JSON output, that makes it usable for real production workflows rather than just chat.

Where the Benchmark Claims Come From

The numbers getting shared around — most notably an 80% Pass@1 score on the DeepSWE coding benchmark — trace back to a 10-task independent evaluation run by a researcher named Ben Davis, not an audited leaderboard. That distinction matters a lot, so it's worth being precise about what's actually being compared:

The catch: DeepSWE and SWE-bench Verified are different benchmarks with different difficulty profiles, so an 80% DeepSWE score isn't directly comparable to a 96% SWE-bench Verified score — despite how the headlines are framing it. It's also a small sample (10 tasks), run by one independent tester, not a controlled leaderboard submission. Treat the "beats GPT-5.6 and Claude" framing as a preliminary signal, not a verified result.

That said, a few specific data points from the same testing are worth flagging on their own:

• On a task called "Meriyah Explicit Resource Declarations," Ox Alpha passed on its first attempt where GLM-5, GPT-5.6 Sol, and Grok 4.6 had all previously scored 0/4.
• It reportedly held a clean pass across 51,469 regression tests during testing, without introducing side effects.
• In a documented 69-tool-call agentic workflow, it made a single error with no retry loops — a decent signal for anyone evaluating it for autonomous agent work specifically.

Who Actually Built It? The Fingerprinting Case

Nobody has officially claimed Ox Alpha. But independent researcher Ben Davis published a technical fingerprinting analysis on August 21 concluding he's about 99% confident it's a variant of Zhipu AI's (Z.ai) unreleased GLM-5.x series. His reasoning rests on four main pillars:

1. Video encoder behavior. Across four controlled video test sets, Ox Alpha's token consumption, frame-rate-independent sampling, duration scaling (~147 tokens/second of video), and per-frame resolution scaling all matched GLM-5V-Turbo almost exactly. Candidate models like MiMo v2.5, Qwen 3.8 Max, and GLM-4.6V all showed clearly different encoding behavior.

2. Tokenizer alignment. Across 25 test prompts, Ox Alpha's token counts matched GLM-5.3 exactly aside from a fixed +75 token wrapper offset — a strong sign of a shared vocabulary.

3. Audio rejection pattern. Ox Alpha rejects audio input the same way GLM-5V does, which also rules out MiMo v2.5 as a candidate, since that model does support audio.

4. Output style. Ox Alpha uses about 1.3 emojis per 1,000 characters of output, consistent with the GLM/Qwen family's style. Claude, GPT-5.6, and Grok all showed near-zero emoji usage under the same test conditions in Davis's analysis.

Organizations his analysis ruled out, and why: Xiaomi (video/audio mismatch), DeepSeek (no prior video capability, different tokenizer), Google, Alibaba, xAI, OpenAI, and Anthropic (tokenizer, output style, or architecture mismatches across the board).

Worth noting: this is inference from behavioral fingerprinting, not a confirmed source. It's compelling circumstantial evidence, not an admission.

This Isn't the First Time — The Stealth Model Pattern

Anonymous "stealth" models testing on routing platforms ahead of an official announcement has become a recognizable pattern, especially among Chinese labs. According to tracking from the platform OrcaRouter, four other stealth models have appeared and later been officially claimed in the past six months:

The timing lines up with Zhipu's own recent moves: the company released GLM-5.3 as a text-only model on August 14, 2026, stating that model weights wouldn't be downloadable until a safety review completed roughly two weeks later. Ox Alpha appeared six days after that, already carrying the multimodal (text, image, video) capabilities GLM-5.3 didn't ship with — which fits the theory that this is a preview of an unreleased, unified GLM vision flagship.

Should You Actually Use It?

If you're evaluating Ox Alpha for real work, here's a grounded way to think about it:

• Good fit for: experimenting with very-long-context workflows (whole-repo analysis, large document review), agentic coding tasks where tool-calling reliability matters, and any workflow that benefits from combining code with visual context (screenshots, UI mockups, video walkthroughs).
• Go in aware of: no official documentation, no support channel, no SLA, and a genuinely temporary free window. There's also no confirmed post-preview pricing, so don't build anything you can't quickly swap out.
• Don't over-index on the benchmark headlines. The 80% DeepSWE number is real, but it's a small, independently-run sample against a different benchmark than the 96%+ SWE-bench Verified scores frontier models report. Run your own tests on your own tasks before drawing conclusions.

Bottom Line

Ox Alpha is a legitimately interesting case study in how frontier AI labs are increasingly using anonymous, free preview releases to gather real production usage data before an official launch — a "testing through distribution" strategy that lets a lab validate infrastructure and gather feedback without committing to a name or a price. Whether or not Zhipu ends up confirming it as GLM-5.x, the pattern itself (four prior stealth models, all later claimed by Chinese labs) suggests this won't stay anonymous for long.

If you're a developer who likes to try new models the moment they show up — the way plenty of us already rotate between GLM, Kimi, and whatever else lands in the OpenRouter list — this is worth 15 minutes before the free window closes around August 27. Just don't build production infrastructure on a model with no name attached to it yet.

Model ID to try: stealth/ox-alpha on OpenRouter.

This piece is a snapshot based on public reporting and independent testing as of August 22–23, 2026. Attribution to Zhipu AI/GLM-5.x is not officially confirmed, and pricing, availability, and benchmark standings can change quickly on a release like this — verify the live OpenRouter listing before relying on anything here for production decisions.