The most-used model on OpenRouter has a name. On August 26, 2026, after six days of anonymous operation as stealth/ox-alpha, Ox Alpha was confirmed as GLM-5.3-Flash, built by Z.ai — the company also known as Zhipu AI. The weights went up on Hugging Face under the MIT licence, and the free preview ended the same day.
The community had it days earlier, from token counts and error codes. What follows is how the six days unfolded, what was actually revealed, what it means for anything you built against the old model ID, and why an open-weights release was the genuinely surprising part.
The six days, in order
| When (UTC) | What happened |
|---|---|
| Aug 20 | stealth/ox-alpha appears on OpenRouter under the anonymous “Stealth” provider: $0 pricing, a 1,048,576-token context window, text/image/video input. No developer named, no model card, no benchmarks. |
| Aug 20–21 | Developers push it through coding agents at scale. Fingerprinting threads begin within hours. Press coverage reports a hosting partner describing a one-week free window. |
| Aug 21 | Ben Davis posts a DeepSWE result: 8 of 10 tasks passed, about 80%, ahead of named commercial models on the same ten. He attaches his own caveat about variance. The caveat does not travel; the number does. |
| Aug 22 | Fingerprinting consolidates around the GLM family. A prediction market on the developer’s identity puts Z.ai in the mid-80s percent. |
| Aug 23 | Davis runs the full 113-task DeepSWE set. The result is roughly 63% — mid-pack rather than table-topping. The correction circulates far less widely than the original. |
| Aug 26, ~09:00 | Bloomberg reports Z.ai as the developer behind the stealth listing. |
| Aug 26, 13:59:01 | The OpenRouter catalogue gains z-ai/glm-5.3-flash. The stealth listing is retired. |
| Aug 26, evening | Z.ai publishes its own announcement, the model card and weights go live on Hugging Face under the MIT licence, and the API documentation appears at docs.z.ai. |
Six days is fast. Previous stealth previews have run longer, and the compressed window is one reason the reveal caught people mid-experiment: a lot of half-finished evaluations were still running against a model ID that stopped existing on a Wednesday afternoon.
The scale of the preview
By the end of its six days, Ox Alpha was the most-used model on OpenRouter — not the most-used stealth model, the most-used model on the platform. A free, anonymous listing with no documentation and no brand had displaced every commercial flagship on the leaderboard, which is roughly what happens when you make a capable long-context model cost nothing.
Precise token volumes are harder to pin down than the ranking, and this is a place where the widely-shared numbers disagree. One figure in circulation is 11.6 trillion tokens, which comes from OpenRouter’s weekly chart; another is 23.2 trillion, a six-day cumulative count. Both are third-party readings of public dashboards rather than anything Z.ai or OpenRouter published, and they are measuring different windows, so treat the order of magnitude as the signal and the precise value as unsettled. We have not seen a primary source that reconciles them, and we are not going to pick whichever is larger.
Either way the point stands: this was not a quiet technical preview. It was one of the largest real-world evaluations any model has had before launch, conducted entirely by people who did not know what they were evaluating.
How the community got there first
The identification did not come from anything the model could do. Capability tells you very little — a multimodal coding model with a long context narrows the field to perhaps a dozen labs and proves nothing. What identified GLM were implementation artefacts nobody thinks to disguise.
The signals that carried the most weight, as reported by the testers who ran them: tokenizer alignment — Ben Davis reported Ox Alpha’s token counts tracking a known GLM model across 25 prompts with a roughly constant offset, and a constant offset is a fingerprint in a way that a coincidence is not; video encoding behaviour matching GLM’s vision models across samples while differing from other candidates; an error code characteristic of GLM’s API surfacing through OpenRouter’s anonymising layer; and the specific way the model rejected audio input, identical in shape to GLM’s vision models. Stylistic tells — emoji frequency, formatting habits — were also cited, though they are much softer evidence on their own.
Those claims were published by the people who ran the tests, chiefly in a fingerprint round-up circulated on August 20 and in Davis’s posts on X, with The Next Web covering the state of the theories on August 22. We reproduced them with attribution rather than reproducing them as fact, and the full ledger — including which signals turned out to be wrong — is on who made Ox Alpha. Two theories lost: Xiaomi’s MiMo line, argued mostly from multimodality and a coding focus, and a Microsoft MAI reading that rested on a single tokenizer analysis contradicting several others.
The lesson generalises, and it is the useful thing to take from this episode: implementation artefacts identify a lab; capabilities do not. The theories that failed reasoned from what the model was good at. The theories that succeeded reasoned from what its plumbing leaked.
What this site got right, and when
Two calls, both made before the reveal, both checkable against our own publication dates.
The identity, on August 22. Our launch coverage reported that “several analyses shared by those testers report a strong match to the GLM family, with the testers themselves quoting confidence figures of roughly 90–99%,” while stating plainly that “none of this is confirmed, no lab has commented, and we treat it as attributed speculation until there is an announcement.” That is four days before Z.ai confirmed it. The point is not that we guessed — the point is that the evidence was reportable without being overclaimed, and it was correct.
The benchmark, on August 23. When the “80%, beats every frontier model” figure was at peak circulation, we published the correction: the 80% came from a ten-task subset, the full 113-task run landed at roughly 63%, and the model was mid-pack rather than ahead of the field. Three days later Z.ai published its own DeepSWE v1.1 figure: 63.4.
A vendor’s launch-day benchmark matching what an independent tester measured blind — before anyone knew whose model it was — is unusually strong corroboration, and it is the single most useful thing to come out of the stealth window. Nobody was tuning a demo, because nobody knew whose reputation was attached. Sites still leading with 80% are quoting a number that was superseded by its own author three days before the model even had a name.
What was actually revealed
| Name | GLM-5.3-Flash — the first natively multimodal model in the GLM-5 series |
| Developer | Z.ai (Zhipu AI) |
| Model ID | z-ai/glm-5.3-flash on OpenRouter · glm-5.3-flash on the Z.ai API |
| Parameters | 320B total, 18B active per token (mixture of experts) |
| Architecture | Hybrid sparse and linear attention; Manifold-Constrained Hyper-Connections; trained on a 30T-token multimodal corpus |
| Context / max output | 1,048,576 / 131,072 tokens |
| Input / output | Text, image, video in; text out. Audio rejected. |
| Weights | MIT licence, zai-org/GLM-5.3-Flash on Hugging Face |
| Technical report | arXiv 2602.15763 — “GLM-5: from Vibe Coding to Agentic Engineering” |
The configuration published alongside the weights fills in what the stealth listing could only imply. The model runs 45 layers, of which 34 use linear attention and 11 use sparse attention in a repeating three-to-one pattern — that ratio is the mechanism behind the million-token context, because it keeps most layers off the quadratic cost curve that makes long context expensive elsewhere. The mixture of experts routes across 288 routed experts plus one shared expert, activating eight per token, which is how 320B total parameters can cost 18B-worth of compute to run. Z.ai also reports a technique it calls Manifold-Constrained Hyper-Connections and a 30-trillion-token multimodal pre-training corpus. We work through what all of that means in practice on the GLM-5.3-Flash reference.
Nothing in the specification changed at the reveal. The context window, the output ceiling and the modalities are all exactly what the stealth listing advertised. What changed was the name, the price, and the fact that you can now read the design decisions instead of guessing at them. Z.ai’s own framing is that the model outperforms the previous generation “at one-tenth the price” while approaching the top of the Claude family on coding and agentic benchmarks — a vendor claim, and one worth checking against independent measurement.
The open-weights release is the real story
Most people expected the stealth preview to end the way they always do: a name, a price, an API. Almost nobody predicted the weights.
Z.ai published GLM-5.3-Flash on Hugging Face under the MIT licence — not a bespoke “open but restricted” community licence with acceptable-use riders and revenue thresholds, but one of the most permissive licences in wide use. Commercial use, modification and redistribution, with essentially no conditions beyond retaining the notice.
Three consequences follow immediately. You can run it yourself, which makes the “should I send my code to a Chinese lab?” question a choice rather than a constraint — we cover the practicalities on self-hosting GLM-5.3-Flash. Anyone else can host it too, and within about a day roughly a dozen providers were serving it on OpenRouter at prices they set themselves. And the model cannot be withdrawn: whatever Z.ai does with pricing or availability later, the weights released on August 26 exist permanently.
That last point is the one worth sitting with. A stealth preview is, by construction, a thing that can vanish. This one ended by becoming permanent.
What changed for you
The model ID is retired, with no redirect
stealth/ox-alpha is delisted and nothing forwards to the new listing. Calls to the old ID fail rather than quietly rerouting — inconvenient for about ten minutes, and much better than the alternative, which would have been silently moving your traffic onto a paid model without telling you.
The migration is a string replacement plus a funded account. Change the ID to z-ai/glm-5.3-flash, then check that the OpenRouter account behind your key has credit, because a zero balance that worked fine during the preview now returns a 402. The base URL, the authentication header, every supported parameter, the context and output limits and the multimodal content format are all unchanged. Full detail, including per-tool configuration, is in the API and migration guide and the coding-tools setup guide.
One thing genuinely worth auditing: anything that looped freely against a $0 model. Retry logic, evaluation harnesses and batch jobs that cost nothing during the preview now have a bill attached. It is a cheap model, but it is not a free one.
It costs money now
| Per 1M tokens | Promotional — until Sept 9, 2026 | List — from Sept 10 |
|---|---|---|
| Input | $0.075 | $0.15 |
| Cached input | $0.015 | $0.03 |
| Output | $0.25 | $0.50 |
The promotion ends at 24:00 UTC+8 on September 9, 2026, at which point the rates double. Note the cached-input line: at one fifth of the normal input rate, it is the difference between a viable million-token workflow and an expensive one, provided you structure prompts so the stable material comes first. Worked examples and the live per-provider price spread are on the pricing page.
You now have three routes, not one
During the preview there was exactly one way to use this model. There are now three: OpenRouter, which routes across a dozen competing providers and fails over when one errors; the Z.ai API directly, using the model ID glm-5.3-flash and documented at docs.z.ai, which cuts out the intermediary; and your own hardware, under the MIT licence. Each makes sense in different circumstances, and the GLM-5.3-Flash reference works through when to pick which.
What this says about the stealth programme
Ox Alpha was the seventh stealth codename since the programme began in spring 2025, and the seventh to be revealed. Not one has ever simply vanished unidentified. The tracker and full reveal history has the complete record.
The pattern is now well enough established to be predictive. A stealth listing appears at $0 with unusually good specifications; developers stress it immediately because it is free; fingerprinting starts within hours; the free window runs somewhere between a few days and a couple of weeks; and it closes with a rename to a priced listing. If you are looking at a live stealth model today, that is roughly the shape of what is coming.
What Ox Alpha adds to the pattern is the demonstration of why labs bother. Six days of genuinely anonymous traffic produced something no internal evaluation can buy: benchmark results from people with no idea whose model they were testing, and therefore no incentive to flatter or to dunk. That the independent number and the vendor number landed within half a point of each other is the strongest possible advertisement for the arrangement.
Frequently asked
Was Ox Alpha GLM-5.3-Flash?
Yes. Ox Alpha was the codename Z.ai used while testing GLM-5.3-Flash anonymously on OpenRouter between August 20 and August 26, 2026. Same model, same specifications — the stealth listing and the released listing differ only in name and price.
Why was Ox Alpha renamed?
It was not renamed so much as unmasked. “Ox Alpha” was never a product name — it was an anonymous codename assigned for a stealth preview, so that Z.ai could collect real-world usage without brand expectations attached. When the preview ended, the model was published under its real name, GLM-5.3-Flash, and the codename was retired.
Is Ox Alpha still free?
No. The free preview ended with the reveal on August 26, 2026. GLM-5.3-Flash costs $0.075 per million input tokens and $0.25 per million output tokens until September 9, 2026, then $0.15 and $0.50. Free routes still exist: run the MIT-licensed weights yourself, or use one of OpenRouter’s rotating free models.
What happened to stealth/ox-alpha?
It was delisted on August 26, 2026 and there is no redirect. Requests to stealth/ox-alpha now fail. Replace it with z-ai/glm-5.3-flash on OpenRouter, or glm-5.3-flash if you are calling Z.ai directly, and make sure the account behind your key has credit.
Sources and verification
Last substantively reviewed: August 27, 2026. Reviewed by: OxAlpha.chat Editorial Team.
How we verified this page
This news article records a dated reveal. The identity and release claims were checked against Z.ai first-party sources, the current OpenRouter listing and the published checkpoint. Later facts are labeled as updates rather than presented as if they were known at first publication.
Primary sources
- Official announcement: Z.ai GLM-5.3-Flash release — confirms the reveal and vendor-reported launch details.
- Official model documentation: Z.ai GLM-5.3-Flash overview — records the model code, Ox Alpha test and capabilities.
- Current platform record: OpenRouter GLM-5.3-Flash listing — provides the current model page and live metadata.
- Published checkpoint: Official Hugging Face model card — confirms the released weights and deployment guidance.
Found a factual error or a changed price? Send us the page URL and supporting source; corrections follow our editorial policy.