Ox Alpha Was GLM-5.3-Flash All Along — Here Is What the Name Changes

GLM-5.3-Flash is the model that spent the last week serving free traffic on OpenRouter as stealth/ox-alpha. Z.ai said so itself on August 26, 2026, and published MIT-licensed weights the same day — which makes the licence the one fact about this model that was genuinely decidable on day one.

The reveal matters most to the people it names last: teams already running the anonymous endpoint in evals or production. For them, nothing about the model's behavior changed at 13:59 UTC — but almost everything about the decision around it did. The provider is now a named company with published terms. The weights are now downloadable under MIT. And the $0 price that made the stealth period so popular now has a dated, three-step exit path.

This post covers what was actually confirmed on August 26, what the named listing says that the archived stealth listing did not, the price path from free to promo to list, the cost-relevant defaults buried in the docs, and an honest read of the vendor's own benchmark table — including where it trails the models it is "approaching."

When the anonymous listing first appeared on August 20, the open question was simply who built it. That question closed on August 26 from both directions at once. Z.ai's launch blog, "GLM-5.3-Flash: Frontier Intelligence, Flash Cost," states it plainly: "Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback."

OpenRouter, for its part, did not quietly delete the stealth slug. The legacy stealth/ox-alpha page stays live as a historical record, now carrying a first-party resolution notice identifying Z.ai as the developer and pointing to GLM-5.3-Flash. The slug itself is gone from the queryable models API: as of August 26, the working IDs are z-ai/glm-5.3-flash for the interactive surface and z-ai/glm-5.3-flash:batch for batch. The named listing entered OpenRouter's public models catalog at 13:59:01 UTC — the same day the bulk of the weight uploads landed on Hugging Face.

The stealth run was not a quiet one. Stripe CEO Patrick Collison called the anonymous model "very impressive" in a post on X, per TechCrunch's August 23 reporting — three days before anyone outside Z.ai knew whose model he was praising. And as OpenRouter's stealth-model track record shows, a reveal was likely but never guaranteed — this one resolved cleanly, with the vendor claiming the model on the record rather than leaving it to leaks and tokenizer forensics.

Almost everything else claimed about GLM-5.3-Flash on August 26 — benchmark standings, serving-hardware efficiency, usage records — needs weeks of independent replication before a careful team should act on it. The licence does not. The Hugging Face repo zai-org/GLM-5.3-Flash carries license: mit in its README front-matter, and the standalone LICENSE file contains the MIT License text with a 2026 Z.AI Co., Ltd copyright line. A second repo, GLM-5.3-Flash-BF16, ships full-precision weights under the same licence.

That is the strongest possible answer to the question a stealth model cannot answer at all. MIT means commercial use, modification, redistribution, and self-hosting with attribution — no use-restriction clauses, no revenue thresholds, no acceptable-use policy layered on top. On this dateline, that puts GLM-5.3-Flash in a different licensing class from the day's other open-weights release: Qwen3.8-Flash-Next ships under the more conditional Qwen Community License 1.0, a contrast we unpack in the Qwen open-weights-versus-hosted-API split. The GLM-5.3 flagship, meanwhile, still has no public weights at all — they were promised, not published, at GLM-5.3's own launch, twelve days earlier — so the Flash's MIT licence stands alone in its family today.

One more thing the card settles: modality. The pipeline tag says text generation, but the tags array includes image-text-to-text and the card body opens by calling this "the first natively multimodal model in the GLM-5 series" — resolve modality against the card body, not the pipeline tag. If you are costing image inputs across providers, our multimodal input billing reference covers how vendors meter them.

"The free ride is over" is not a metaphor here. The August 24 Wayback capture of the stealth listing states its price in three words — "free to use" — so free was the listing's own stated price throughout the preview, not a retrospective characterization. Free was step one. The named endpoint adds two further steps and one date.

Step two is live now: OpenRouter's models API prints the interactive z-ai/glm-5.3-flash surface at $0.075 per million input tokens and $0.25 per million output tokens, with cache reads at $0.015/M. Step three is reported for September 9: per Z.ai's framing as relayed by two independent secondary outlets, the launch promo — 50% off — runs through September 9, 2026, 24:00 UTC+8, after which the interactive surface reverts to list at $0.15/M in and $0.50/M out. We did not find that date stated verbatim on a Z.ai pricing page in this research pass, so treat it as reported rather than as a direct vendor quote — but the mechanism behind it is primary-confirmed, because the :batch surface already prints the $0.15/$0.50 list rate today.

That makes for a genuinely odd dateline fact: right now, batch is the expensive surface — exactly double the interactive promo on every axis, cache reads included. Once the promo lapses, the two surfaces print the same numbers as listed today. Whether a batch surface earns its keep in your stack is a separate decision with its own math — we covered the batch-pricing landscape generally earlier this month — but on this model, on this date, "batch is cheaper" is simply false.

One adjacent perk for subscribers rather than API callers: Z.ai's docs state that GLM Coding Plan subscribers get 3x the usable quota when routing to GLM-5.3-Flash versus GLM-5.3, on a points system where off-peak and weekend calls cost half the standard points. That is a plan-economics detail, not an API price — but it signals how aggressively Z.ai wants Flash to absorb coding traffic.

Most coverage of the reveal describes the named listing as if it had been static all along. It was not — and the cleanest way to see that is to diff the archived stealth listing against the live post-reveal API. The Wayback Machine's August 24, 2026 capture of the stealth listing states, in the listing's own copy: "free to use. 1,048,576 token context window, maximum output of 131,072 tokens."

The named interactive listing now lists 1,310,720 tokens of context — 25% more than the archived stealth figure — while :batch lists 1,048,575, effectively the stealth-era cap. Maximum output stayed at 131,072 on both surfaces. To be precise about what the evidence supports: the Aug-24 capture shows the smaller number, and the live August 26 API shows the larger one on the interactive surface. No dated capture pins the change to reveal day itself, and no reason for the delta appears in any source we checked — the Z.ai launch blog, the docs page, or OpenRouter's models API record and legacy stealth listing. A deliberate capacity bump for the named launch and a batch surface provisioned from an older config snapshot are both live explanations — we are reporting the difference, not asserting the intent.

The terms row deserves a beat. The archived stealth listing stated: "Prompts and completions are retained by the provider and are not used for training; all other use is governed by the Stealth Model Terms." That promise was attached to the stealth listing — it does not automatically travel to the named endpoint, and the reveal is precisely the moment to re-check which retention and training terms now govern your traffic. Where free and promo tiers across the industry actually land on that question is the subject of our census of free and promo-tier data terms.

Two lines in Z.ai's docs matter more to your invoice than any benchmark. First: thinking.type "only supports enabled" — thinking cannot be disabled on this model, the same mandatory-reasoning pattern we reported when the flagship's API terms made thinking non-optional there. Second: reasoning_effort accepts exactly three values — low, high, and max — and defaults to max if you set nothing. Every unconfigured call runs the most expensive configuration the API offers. With output on the interactive surface billed at $0.25/M now and $0.50/M at list, and reasoning tokens billed as output, an unexamined default compounds quietly.

The docs' recommended inference settings are temperature: 1 and top_p: 0.95. What Z.ai does not publish anywhere we checked — the launch blog, the docs page, and the Hugging Face card — is any RPM or TPM rate limit for the GLM-5.3-Flash API. That is a confirmed absence in those sources, not an unfound number: if your production plan needs a throughput ceiling, you will have to establish it empirically or ask Z.ai, because no published figure exists to cite.

All of this belongs to a broader discipline: a reveal is not reassurance, it is a re-verification trigger. We keep the full list of what to re-check the day an anonymous model gets a name — terms, pricing surfaces, caps, defaults, identity continuity — in the stealth-model reveal-day checklist.

The alias point is worth its own sentence: the floating alias that already points at a GLM-5 model is exactly the mechanism by which a team that thinks it is calling one model ends up calling another after a release week like this one. If your routing configs, evals, or cost dashboards reference anything other than the dated slug, this is the week to fix that. For teams deciding whether an MIT-licensed 18B-active model earns a lane in their routing mix at all, that comparative evaluation is the kind of work our AI transformation engagements start with.

Z.ai's launch blog publishes a comparison table, and its own prose characterizes GLM-5.3-Flash as approaching Claude Opus 4.8 — not beating it. The table bears that out. On Terminal-Bench 2.1, Flash scores 84.3 against Opus 4.8's 85.0 and GPT-5.6 Terra's 87.4 — it trails both. On individual sub-benchmarks the picture genuinely splits: Flash scores higher than Opus 4.8 on DeepSWE v1.1 (63.4 vs 58.0), AutomationBench v1.0.6 (48.8 vs 41.0), Toolathlon Verified (78.4 vs 76.2), GDPval-AA v2 (1773 vs 1582), OfficeQA Pro (62.4 vs 48.9), and Chartography with Tools (78.0 vs 75.0) — while trailing on Terminal-Bench 2.1, NL2Repo, HLE with Tools, CharXiv, BabyVision, MVbench, and MMVU. Beating Opus 4.8 on roughly half the published metrics and trailing on the other half is a remarkable result for an 18B-active model at this price — but "beats Opus 4.8" would be the wrong summary, and Z.ai itself does not make it.

The launch blog makes one more claim worth quoting exactly, because it is the kind that usually arrives as rumor and here arrives on the record. Of the stealth run, Z.ai writes: "It quickly became the most popular model of the week — with all of this traffic served on Chinese AI chips." And on efficiency: "Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale."

Treat both statements as what they are: vendor-stated, on the record, and dated August 26. Two honest caveats attach. Z.ai names no chip vendor, and it publishes no independent throughput or power figures — so as of the reveal, no named third party had confirmed it in the sources we checked. If accurate, a week of frontier-scale free traffic served on domestic silicon at claimed NVIDIA-comparable per-token cost would be a significant infrastructure datapoint on its own. That "if" is doing real work, and it is Z.ai's to discharge.

How much traffic did the free week actually serve? Here is the uncomfortable answer: the only usage figure in this story that traces cleanly to a primary source is a single-day run rate. OpenRouter posted it from its official X account on August 24.

That is one day, dated, from the platform operating the endpoint — and it is where the solid ground ends. Cumulative totals for the stealth period circulate widely in secondary coverage, and they disagree with each other by roughly a factor of five, with none of them traceable to a primary source. Z.ai's own launch blog contains zero token-count figures — it says only that the model became "the most popular model of the week" — so any period total attributed to Z.ai is a misattribution by construction. Some of the circulating numbers appear to conflate the 30-trillion-token pretraining corpus with inference volume, which are different quantities measuring different things. That fivefold disagreement is itself the finding: chart-derived usage numbers get re-reported without their definitions, windows, or provenance, and mutate along the way. What OpenRouter's usage charts actually measure — and why totals read off them so rarely survive re-derivation — is exactly the subject of our companion piece on OpenRouter's usage charts.

Our projection from here is straightforward. The named endpoint now has real prices, so the stealth-era volume — whatever its true total was — becomes the wrong baseline for forecasting paid demand: free-tier traffic measures curiosity, not willingness to pay. The more durable signal is structural: a vendor was willing to spend a week of frontier-scale serving costs to seed an audience before attaching a name, a licence, and a price. Given how well the sequence appears to have worked — an audience, a Stripe-CEO endorsement, and a same-day MIT release to convert goodwill into deployments — expect the anonymous-preview playbook to be reused, and expect the usage-total fog to be a recurring feature of reveal weeks, not a one-off.