What is Ox Alpha? Stealth model revealed as GLM-5.3-Flash

Last updated: August 27, 2026. Update log: Aug 27, Z.ai confirmed Ox Alpha as GLM-5.3-Flash and released the weights, article updated throughout. Aug 25: original publication, identity still unknown at the time.

On August 20, 2026 a new model appeared on OpenRouter under the ID stealth/ox-alpha: a million-token context window, image and video input, a $0 price tag and no lab name attached. For six days it swallowed trillions of tokens from coding agents, earned a "very impressive" from Stripe CEO Patrick Collison and set off the best guessing game the AI community has had all year. Then the mystery ended.

We published the first version of this article two days before the reveal, when the strongest available evidence was a tokenizer fingerprint. That evidence turned out to be exactly right. Here's the full story, updated: what was confirmed, what the stealth week looked like from the inside and what changes now that the model has a name.

Ox Alpha revealed as GLM-5.3-Flash

On the evening of August 26, Z.ai (Zhipu) confirmed that Ox Alpha was GLM-5.3-Flash all along and released the weights on Hugging Face under an MIT license. The announcement filled in the numbers the stealth label had been hiding: 320B total parameters with 18B active (a sparse MoE design), the same 1M-token context window and native multimodal input. The part that made mainstream headlines is the hardware. Z.ai says the entire stealth trial ran on a cluster of 100,000 domestically produced Chinese chips, and Zhipu's shares closed 12% higher in Hong Kong the day after.

The stealth run itself set records. The model processed 62 trillion tokens across OpenRouter and OpenCode during the preview, OpenRouter's biggest launch to date, and at its peak it ranked first for coding with roughly 31% of the platform's weekly volume. The named API is priced at $0.15 per million input tokens and $0.50 per million output.

Ox Alpha specs and pricing

These are the specs as they stood during the preview, from the OpenRouter listing, which was the model's only official surface for six days:

• Model ID: stealth/ox-alpha, served by a single provider labeled "Stealth"
• Context window: 1,048,576 tokens, with a 131,072-token max output
• Input: text, images and video; output is text only
• Pricing: $0 per million tokens during the preview; the named GLM-5.3-Flash API now runs $0.15 in and $0.50 out
• Features: tool calling and JSON output via response_format; a reasoning_effort parameter (low, high, max) that defaults to max

The listing described a reasoning model aimed at coding and sustained agentic work, and it was blunt about the mystery: the model "is developed and operated by a third-party provider who has chosen to remain anonymous during this preview". Z.ai's reveal post called the stealth listing a preview skin, so expect the stealth/ox-alpha endpoint to be retired and plan any integration around the named model instead.

Who is behind Ox Alpha? Z.ai, and the fingerprints called it

The reveal confirmed what the technical evidence had been saying since the first 48 hours. This section stays in the article deliberately, because it's a nice case study in how model attribution works when nobody will tell you anything.

The GLM-5.3 fingerprint evidence

The tokenizer test was the strongest signal. Probe prompts mixing languages, emoji, code and SQL produced token counts that matched GLM-5.3 on every single probe, 11 out of 11 in the set collected by DigitalApplied's stealth model census, while no other lab's model matched more than four. A tokenizer is about as close to a fingerprint as this field gets, and this one didn't lie.

The serving stack agreed. A leaked stack trace from a malformed request contained a Java class path matching Zhipu's documented cloud API internals, the error codes on bad requests followed the GLM API's dialect and a separate analysis found Ox Alpha consuming video tokens the same way GLM-5V-Turbo does. Manifold bettors had priced Zhipu at roughly 90% days before the confirmation. The market was right too.

The Gemini and Microsoft MAI theories

Early chatter guessed Google, mostly on timing and feel, and TechCrunch's report mentioned a theory about an unreleased Microsoft MAI model. Neither survived contact with the fingerprint work, and neither turned out to be true. Worth remembering the next time a stealth model drops: output "vibes" identified nothing, infrastructure forensics identified everything.

Ox Alpha benchmark results: the 80% claim and the real score

Here's where the hype ran ahead of the data during the stealth week, and where a lot of the day-one coverage aged badly.

The number that went viral was an 80% Pass@1 on DeepSWE, from a 10-task sample someone ran on launch day. Ten tasks. A coin-flip-sized sample, repeated across half the internet as "mystery model beats every frontier lab".

The full run told a calmer story. StealthModelWatch's complete 113-task DeepSWE run scored 58.4% Pass@1 (66 of 113 solved, 95% CI of 49.2% to 67.1%), roughly level with the top commercial coding models. Their report was direct about the earlier figure: "the rumored ~80% pass rate is completely incorrect." Z.ai's own launch materials now cite 63.4 on DeepSWE v1.1, in the same neighborhood as the independent run and nowhere near the viral number. The community measurement held up well.

My read stands from the first version of this article: frontier-class at coding, without being the leaderboard anomaly the early posts described. And a 58% from an anonymous free preview was the impressive number all along. It didn't need inflating.

Ox Alpha speed and latency

The consistent complaint during the preview was speed. Public telemetry hovered in the mid-20s tokens per second at the median, with a few seconds before the first token and P95 tails that stretched much longer at peak hours. Reasoning runs before every answer and defaults to max effort, so anything interactive felt sluggish. Now that the load spreads across the named API and self-hosted deployments, expect this to improve, though the reasoning-first design means it will never feel like a lightweight chat model.

Ox Alpha data retention and the stealth model terms

This mattered during the preview and the lesson outlives it.

OpenRouter's stealth program has a default deal: the model is free and the anonymous provider can collect your prompts and train on them. That's the exchange, spelled out in the Stealth Model Terms. Ox Alpha carried an exception, the only one among all 14 stealth models to date: prompts and completions "are retained by the provider and are not used for training".

Everything sent during the stealth week now sits, retained, with a named company: Zhipu. That's better than an anonymous one, and it changes nothing about what already left your machine. Working at a hosting provider means I read data-handling clauses for a living, and the rule I gave here on day one is the one I'd repeat for the next stealth model: treat every prompt as public. Throwaway branches and open-source code are fine. Proprietary code and customer data are not.

Is oxalpha.com the official Ox Alpha website?

No, and it never was. Neither oxalpha.com nor oxalpha.io belongs to Z.ai. Both are third-party front-ends that appeared within days of the launch to catch the search traffic, and to their credit both said so. Now that the model has a name, the official surfaces are Z.ai's own site and the zai-org account on Hugging Face, where the GLM-5.3-Flash weights live. Treat any oxalpha-something domain as an unofficial wrapper that adds its own terms and its own logging between you and the model.

How to access GLM-5.3-Flash now

The picture after the reveal, in order of practicality. The named API from Z.ai runs $0.15 per million input tokens and $0.50 out, with cached input at $0.03. OpenRouter will carry the named model as it did the stealth one, so agent harnesses that pointed at stealth/ox-alpha need a one-line model-ID change. And since the weights are MIT-licensed on Hugging Face, big teams can run it on their own hardware, with the caveat that 320B total parameters is multi-GPU territory, well beyond a hobbyist card. We went over the verified benchmarks and the real hardware requirements in our GLM-5.3-Flash article, so start there before planning anything.  

The free ride is over either way. If you're hunting no-cost options for agent setups, our guide to free AI models covers the ones that stay free after their hype week.

How the Ox Alpha reveal fits the stealth model pattern

Six days from listing to confirmation. That fits the historical window almost exactly: since April 2025, OpenRouter has listed 14 stealth models and half were later claimed officially. Quasar Alpha and Optimus Alpha ran under two weeks before OpenAI confirmed both as GPT-4.1 test builds, and Horizon Alpha lasted about a week before turning out to be an early checkpoint of the GPT-5 family, the same family we covered in our GPT-5 guide last summer.

Names move fast in this space regardless. We traced OpenClaw's two rebrands in a single month once, and that story still gets referenced today for the same reason this one will: somebody has to write the timeline down while the links are live.

The closing thought from the first version of this article held up, so it stays with one word changed. A Chinese lab put a frontier coding model in front of a few hundred thousand Western developers for a free week, without saying a single word about itself, then open-sourced it. That silence outperformed every launch post I've read this year.