Ox Alpha: The Anonymous 1M-Context Coding Model That Has Everyone Guessing Z.ai

Ox Alpha is a free, anonymous "stealth" reasoning model that appeared on OpenRouter on August 20, 2026, and within 48 hours it had the entire AI community doing forensics. No company has claimed it, no official leaderboard lists it, and yet by August 23 it was already the second-most-used model on the competing OpenCode routing platform, carrying real traffic from Claude Code and Nous Research's Hermes Agent. The mystery matters because everything interesting about Ox Alpha runs through it: a 1M-token context window free for one week, three modalities at once, and an independent researcher who says he is "99% certain" the anonymous weight is actually Zhipu AI's unreleased GLM-5.x flagship testing itself in public before it has a name.

What Ox Alpha Is

The model card on OpenRouter describes Ox Alpha as "a reasoning model designed for coding, sustained agentic work, and production workloads." That is a specific positioning, and it is worth taking literally rather than as marketing: it is aimed at long-horizon software engineering, complex reasoning, and workflows that mix text with visual context. This is not a general chat model; it is built for the kind of long-running agent loops and large-repository work that burns context windows by the hour.

The Spec Sheet

OpenRouter's own listing is the only authoritative source, and it pins the model down precisely.

• Model ID: stealth/ox-alpha
• Released: August 20, 2026
• Price: Free ($0 per million tokens in and out)
• Context window: 1,048,576 tokens (1M)
• Maximum output: 131,072 tokens
• Input modalities: Text, image, and video
• Output: Text
• Tool calling: Supported
• Reasoning: Supported
• Uptime (3-day): 99.99%

The 1M context window is the giveaway of the model's intent — it is enough room for a multi-hour agent session or an entire repository. The 131K output cap allows long single generations. Tool and function calling plus structured JSON output round out the checklist that real agent harnesses require.

Who Is Already Using It

Adoption was the first story. OpenRouter publishes app-level token share per model, and on Ox Alpha's opening days the leaderboard was dominated by agent harnesses rather than chat apps. Claude Code burned roughly 9.32 billion tokens on the model in its first days, and Hermes Agent — Nous Research's persistent open-source agent — drove roughly 8.98 billion more. Oh-My-Pi, DeepSeek Harness, and Z Code also appeared among the top senders. On OpenCode, the model raced to No. 2 within three days, processing about 12 trillion tokens across 180,000 unique users and 3.56 million sessions by August 23.

That traffic is a behavioral signal more persuasive than any single benchmark. Teams running real coding agents were routing heavy, production-shaped workloads to Ox Alpha almost immediately — a model empty of hype and reputation still pulled agent traffic at frontier scale.

The Big Question: Who Built Ox Alpha?

Ox Alpha is what OpenRouter classifies as a "stealth" model: a preview release from a third-party provider who has chosen to remain anonymous during the trial period. OpenRouter routes requests to it but is explicit that it is not the developer, owner, or provider of the weights. Under those terms, prompts and completions are retained by the provider but not used for training; all other use falls under OpenRouter's Stealth Model Terms.

Anonymous releases are a recognized pattern on OpenRouter. Models ship under code names like Hunter Alpha, Healer Alpha, and Owl Alpha; historically, the providers either claim them after the trial or let them fade, and Ox Alpha has so far been claimed by no one. That anonymity is exactly what has set the forensics community to work.

The Benchmark Picture: Read the Fine Print

The benchmark story around Ox Alpha is a lesson in sample size, and the community learned it in real time. The headline figure that swept social media — an 80% Pass@1 score on DeepSWE beating Claude Fable 5 and GPT-5.6 Sol — came from a 10-task subset of the benchmark. Developer Ben Davis ran that evaluation and reported 80%. But DeepSWE itself consists of 113 tasks written from scratch across 91 active open-source repositories in five languages. Ten tasks is under 9% of the set, and in a set that small, a single task shifts the score by roughly ten percentage points.

10-Task Sample vs. Full 113-Task Run

A separate, independent full-set evaluation landed at roughly 63% — level with GPT-5.6 Sol mid and below Claude Fable 5's 65%. The corrected reading is blunt: on a 10-task slice Ox Alpha looked class-leading; on the full independent set it is a good model, not a frontier-beating one.

What matters is the keyword: independent. Only OpenRouter's own listing is authoritative; every benchmark figure above is a community run or a single developer's eval — useful signal, not a verified leaderboard, and all of it is the body of evidence the mystery is built on.

Where Ox Alpha Actually Shined

Benchmarks aside, independent testing turned up specific strengths. Ox Alpha passed the Meriyah explicit resource declarations task on its first attempt, where GLM-5, GPT-5.6 Sol, and Grok 4.6 had all recorded 0/4 results. It maintained a clean pass across all 51,469 regression tests, indicating code generation that does not introduce unintended side effects. And in a documented agent workflow involving 69 tool calls, the model made only one error with no retry loops and low inference overhead — a real signal for long-horizon agent work.

On Kingbench — a separate reasoning benchmark — Ox Alpha scored 87.5, second only to GLM-5.3's 91.25 and ahead of Opus 4.8 (80%). That is a notably strong reasoning showing for a model with no official name.

The Fingerprint: Zhipu AI's GLM-5.x

The identity of Ox Alpha's maker is the crux of the story, and every independent trace leads the same direction: toward Zhipu AI, the Beijing lab behind the GLM family. No one at Z.ai has confirmed anything, but developer Ben Davis says his fingerprinting gives him about a 90%+ confidence the model is a GLM-5.3-series variation, while others read it as an unreleased GLM-5.x flagship. The community consensus leans strongly toward "GLM."

Why the Forensics Point to GLM

Four independent tracks make up the evidence:

• Video encoder behavior: Across multiple controlled video tests, Ox Alpha consumed video tokens in a pattern almost identical to GLM-5V-Turbo's, including frame-rate-independent frame sampling and a matching token-per-second of video ratio. Competing candidates — MiMo v2.5, Qwen 3.8 Max, GLM-4.6V — all behaved differently.
• Tokenizer alignment: Across controlled prompts, Ox Alpha's token counts matched GLM-5.3 almost exactly, differing only by a fixed wrapper — implying a shared vocabulary.
• Audio-interface behavior: Ox Alpha rejects audio input in a manner consistent with the GLM family, and differs from audio-input-capable models like MiMo v2.5 — weakening alternate attribution.
• Output style: Ox Alpha's output style aligns closest with the GLM/Qwen family relative to the other competitors tested.

Additional forensics reported on August 22 — a stack trace, error code 1214, and a 30/30 tokenizer match — all converge on the same GLM family. A Manifold prediction market on the model's origin lists Z.ai/GLM as the leading contract. Analysts like teortaxesTex go further, suggesting a further-trained Ox Alpha could be the workhorse behind a much stronger GLM-5.5.

Why Anonymous, and What "Free for a Week" Means

The anonymity may be deliberate strategy. Anonymous "stealth" releases are how a lab tests frontier-scale real-world traffic without putting its name on the door. If the GLM reading is right, the claimed 100 trillion tokens per day of serving capacity becomes easier to square — Zhipu is one of the few operators with infrastructure that could back that promise. It is reading less like a spec and more like a stress-test challenge: a lab shipping a free anonymous model wants a wave of frontier-scale usage to observe its model under real load.

Data Handling: Retention vs. Training

The most practical distinction on the OpenRouter route concerns data: the provider retains prompts and completions but does not use them for training. On the OpenCode route, the model ships with zero data retention. For anyone pushing proprietary code through the model, that differs matters more than the benchmarks: an anonymous provider with retained logs is a different risk profile than Anthropic or OpenAI under a signed enterprise agreement, even when the sticker price is zero.

Pricing After the Free Week

The free period ends around August 27, 2026 — and no rate card has been published. Historically, stealth previews either graduate to a named model with list pricing or silently disapphese. Anyone building around Ox Alpha should treat "free" as a preview subsidy, not a permanent tier, and hold fallback routes ready.

The Takeaway

Ox Alpha is the best live demonstration in recent memory of how frontier AI actually gets tested in 2026. A model appears overnight with no name, no bench, and no blog post, and inside a week it accumulates hundreds of thousands of users, shows coding strengths that established frontier labs miss — the 51,469 clean regression tests, the single-error 69-tool-call agent run — and earns a credible community identification as one of the biggest Chinese labs' newest flagships. Everything about it — adoption, benchmarks, evidence of identity, pricing, transparency, data retention — has been running in public, and the public has been grading.

The honest reading, for now, is that Ox Alpha is good, not great: at full-set DeepSWE it sits with GPT-5.6 Sol mid and just below Claude Fable 5. But it is free for now, multimodal, and strongly suspected to be a Zhipu AI GLM family model — a combination that puts genuine pressure on the US frontier labs even before the model has a name, a price, or a claimed home. When the hood finally comes off, the only question will be whether the flagship that showed the world its best work gets to introduce an even better model under its own name.

---

Sources: OpenRouter Ox Alpha · OrcaRouter analysis · Local AI Zone deep-dive · explainx.ai · North Denver Tribune