Ox Alpha Was GLM-5.3-Flash: Specs, Price, Access

For six days in August 2026, the most-used model on OpenRouter had no name attached to it. stealth/ox-alpha appeared on 20 August at $0 per token with a 1M-token context window and no lab willing to claim it. On 26 August 2026 Z.ai confirmed it was GLM-5.3-Flash, released the weights on Hugging Face under the MIT licence, and switched on paid pricing on a real model ID.

This page is the settled record of what Ox Alpha was, what the model actually is, and how to use it now. Everything below was re-verified on 31 August 2026 against OpenRouter's live model API, the Hugging Face model card, Z.ai's pricing documentation, Ollama's model library and Artificial Analysis.

What was Ox Alpha?

Ox Alpha was the anonymous preview name for GLM-5.3-Flash, a model built by Z.ai (formerly Zhipu AI), the lab behind the GLM family. It ran on OpenRouter as stealth/ox-alpha from 20 to 26 August 2026 at zero cost, with the operator's identity withheld for the duration of the test.

OpenRouter's model page now carries the confirmation directly. The banner on the archived Ox Alpha listing reads: "This stealth model was developed and operated by ZAI, revealed to be ZAI GLM-5.3-Flash." The page's FAQ repeats it — "Ox Alpha was revealed to be ZAI GLM-5.3-Flash, which can be used here: https://openrouter.ai/z-ai/glm-5.3-flash" — and points at the live listing.

Z.ai says the same thing in its own distribution channels. The Ollama model library entry states plainly: "Before release, Z.ai tested GLM-5.3-Flash anonymously as ox-alpha where it became the most popular model of the week."

The preview is over, and the endpoint is gone. We pulled OpenRouter's public /api/v1/models endpoint on 31 August 2026: 396 models listed, zero entries under the stealth/ namespace. The descriptive page at openrouter.ai/stealth/ox-alpha still resolves as an archive, but the model ID is no longer in the callable catalogue. Any code still sending "model": "stealth/ox-alpha" needs updating.

How was Ox Alpha revealed, and what was the evidence?

Z.ai made the reveal itself on 26 August 2026, publishing GLM-5.3-Flash alongside open weights rather than leaking it. OpenRouter created the z-ai/glm-5.3-flash listing at 13:59 UTC on 26 August 2026 (from the created timestamp in its models API), and the Hugging Face repository zai-org/GLM-5.3-Flash went public the same day.

By the time the announcement landed, the community had already narrowed it down. Independent developers had matched Ox Alpha's tokenizer output to GLM-5.3 across dozens of probe prompts, matched its video-token consumption to GLM-5V-Turbo, and noted GLM-characteristic audio rejection and error surfaces. None of that was proof — but all of it pointed at the same lab.

What our own fingerprint check got right, and what it missed

We ran a different test on 23 August and published it here: instead of probing model behaviour, we diffed the configuration metadata OpenRouter exposes on its public models endpoint. stealth/ox-alpha and z-ai/glm-5.3 returned an identical parameter surface — the same supported parameters, the same defaults (temperature: 1, top_p: 0.95), the same context and output ceilings, and the same unusual mandatory-reasoning block with efforts low/high/max defaulting to max. z-ai/glm-5.2 differed on all of it. Our conclusion was that the fingerprint was specific to the GLM-5.3 generation, not to Z.ai generally.

Being precise about the scorecard, because that is the only version worth reading:

• Right: the lab (Z.ai) and the generation (GLM-5.3). The reveal confirmed both.
• Incomplete: we said Ox Alpha had to be an unreleased multimodal sibling of GLM-5.3, since GLM-5.3 is text-only. That framing was correct but the community names in circulation — "GLM-5.3V", "GLM-5.5" — were wrong. It shipped as GLM-5.3-Flash, a separately trained base model rather than a vision head bolted onto GLM-5.3.
• Wrong: we suggested that if the attribution held, the natural post-preview price point was GLM-5.3's $1.40 / $4.40 per million tokens. GLM-5.3-Flash launched at roughly one-ninth of that. Anyone who sized their budget off our estimate got a pleasant surprise, but the estimate was still wrong.
• Unknown to us: the 320B/18B architecture, the hybrid sparse-plus-linear attention design, and the MIT open-weights release. Metadata fingerprinting tells you which serving stack a model runs on. It tells you nothing about what is inside it.

One postscript worth recording, because it only became visible after launch: the stealth endpoint was running a restricted configuration. Re-pulling the models API on 31 August, z-ai/glm-5.3-flash exposes twenty parameters including structured_outputs, seed, logprobs, stop and the penalty family — all of which the stealth listing withheld. The "no JSON-schema enforcement" limitation everyone attributed to the model during the preview was a property of the preview endpoint, not of GLM-5.3-Flash. Support still varies by provider on an aggregated listing, so check the provider you route to.

What is GLM-5.3-Flash?

Z.ai's own model card describes it as "the first natively multimodal model in the GLM-5 series" — not a multimodal adapter on an existing checkpoint, but a newly trained base model. Verified specifications:

Two architectural details are doing the real work behind the price. The hybrid attention stack mixes sparse and linear attention specifically to cut long-context serving cost, which is what makes a 1M window economically viable at Flash pricing. And an 18B active path out of 320B total means each token only pays for a fraction of the model — the standard MoE trade, executed at an unusually aggressive ratio.

Z.ai's summary of the result: GLM-5.3-Flash "outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks."

How do you use GLM-5.3-Flash now that the stealth listing is gone?

Four routes, in descending order of how quickly you can be running.

Pricing first, because it changed the calculus

Source: Z.ai's official pricing page. The launch promotion is a flat 50% discount running to 24:00 on 9 September 2026 (UTC+8). OpenRouter's models API on 31 August returned the promotional rates for z-ai/glm-5.3-flash ($0.075 in / $0.25 out / $0.015 cached), so the discount is being passed through there too. Budget against the $0.15 / $0.50 list price, not the promo — you have days, not months.

The cached-input rate is the number to design around. At $0.03 per million on list pricing, an agent loop that resends the same system prompt and repository context on every turn costs a fifth of what the uncached path does.

1. OpenRouter

Change the model string and nothing else. Twenty-one providers serve it, plus a z-ai/glm-5.3-flash:batch variant for asynchronous jobs.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [
      {"role": "user", "content": "Refactor this module and explain the tradeoffs."}
    ],
    "reasoning": {"effort": "high"},
    "max_tokens": 8192
  }'

2. Z.ai's own API

Z.ai exposes an OpenAI-compatible surface at https://api.z.ai/api/paas/v4/ with the model ID glm-5.3-flash. Point any OpenAI SDK at it:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.z.ai/api/paas/v4/",
    api_key=os.environ["ZAI_API_KEY"],
)

resp = client.chat.completions.create(
    model="glm-5.3-flash",
    messages=[{"role": "user", "content": "Summarise this repo's failure modes."}],
    reasoning_effort="high",
    max_tokens=8192,
)
print(resp.choices[0].message.content)

Note reasoning_effort. Z.ai's model card is explicit that it defaults to max if you pass nothing — or if you pass anything it does not recognise. Leaving it unset is the single most common way to burn output tokens on a model whose whole appeal is cost. Use max only when you are reproducing benchmark numbers.

3. Ollama — cloud only, not local

A lot of people searching for "ox alpha ollama" want the local route. Be clear on what is actually available: Ollama's library lists one tag, glm-5.3-flash:cloud. It runs on Ollama's hosted infrastructure, not on your machine.

ollama run glm-5.3-flash:cloud

Ollama also ships one-line launchers into agent harnesses — ollama launch claude --model glm-5.3-flash:cloud, and the same pattern for OpenCode. Ollama's listing states the cloud model is hosted in the United States and Europe with zero data retention, which is a materially better data posture than the stealth preview offered.

4. Open weights, run it yourself

This is the part the stealth preview could never give you. The weights are on Hugging Face at zai-org/GLM-5.3-Flash under the MIT licence — the permissive one, with no custom commercial-use rider. The repository had passed 379,000 downloads by 31 August 2026. Z.ai lists supported serving stacks as SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, and a BF16 repository ships alongside the FP8 default.

The honest hardware caveat: 320B total parameters is 320B total parameters regardless of how few are active. Even at 4-bit quantisation that is roughly 160 GB of weights before any KV cache, so this is multi-GPU or high-memory-workstation territory, not a laptop model. Community quantisations in GGUF, NVFP4, MLX and EXL3 formats appeared within days if you want to trade quality for fit — our guide to running GLM-5.3-Flash locally works through the hardware maths, and the GLM-5.2 local setup guide covers the same stack for the smaller sibling.

Was Ox Alpha actually as good as the benchmark hype said?

During the preview, one number travelled further than every other fact about the model: 80% on DeepSWE, beating GPT-5.6. That figure came from a single independent developer running a 10-task subset of DeepSWE and passing 8 of them. It was never an audited leaderboard entry, and the comparison baselines quoted next to it came from different, larger runs — on a 10-item test every score has to be a multiple of ten, so the reported 65% and 62% for other models plainly were not measured on the same sample.

Now there are vendor numbers to check it against. Z.ai's own published DeepSWE v1.1 result for GLM-5.3-Flash is 63.4. The viral community figure overstated the model by roughly 17 points, which is exactly what you would expect from an 8-of-10 sample: one extra pass or fail moves the score ten points.

Here is what Z.ai actually published with the release:

These are vendor-published, self-selected benchmarks — treat them as a claim, not an audit. But they are consistent with the independent picture: Artificial Analysis, which did not list the model at all during the preview, now scores GLM-5.3-Flash at 57 on its Intelligence Index with a measured output speed of 43.8 tokens per second — well above the median of 29 for open-weight models of comparable size.

The fair reading: Ox Alpha was not the GPT-5.6-beating monster the viral number implied. It is a strong agentic-coding model that lands within a point or two of Claude Opus 4.8 on Terminal-Bench and Z.ai's own code benchmark, while costing roughly one-ninth of GLM-5.3 and a small fraction of frontier closed models. The story was never that it won on capability. It was that it got that close on price. For a like-for-like reference across consistent suites, our Kimi K3 benchmark comparison runs the same models through matched benchmarks, and the full GLM-5.3-Flash guide goes deeper on capability-by-task.

What does the stealth launch tell you about how labs ship models?

This is the durable lesson, and it will outlast the news cycle. Anonymous previews on OpenRouter are now a standard release channel, not a stunt. The precedents are consistent: Quasar Alpha and Optimus Alpha (April 2025) turned out to be OpenAI's GPT-4.1; Sonoma Sky Alpha and Sonoma Dusk Alpha (September 2025) turned out to be xAI's Grok 4 Fast. Ox Alpha is the third instance of the same playbook in eighteen months, and the most successful.

What the lab buys with a week of free tokens:

• Unbranded evaluation. Developers judged Ox Alpha's output with no logo attached — no China-lab discount, no frontier-lab halo, no prior. For a Chinese lab competing against incumbent brand trust, that is the single most valuable thing the format offers.
• Production-shaped load testing. Free frontier inference attracts exactly the traffic internal evals cannot reproduce: hour-long agent loops, malformed tool schemas, adversarial prompts, 900K-token contexts. Six days of that is a better stress test than any staging environment.
• Distribution telemetry. Real prompts across real codebases are the hardest training and evaluation signal to buy, and a coding model needs precisely that distribution.
• A launch narrative that costs nothing. A week of "who made this?" delivered an audience already convinced of the model's quality before the price was announced. By reveal day the pitch was not "try our model" but "the model you already liked is ours, here are the weights, here is a price nine times below the alternative."

The pattern also tells you what to expect operationally. A stealth listing has no SLA, no deprecation policy and no notice period — Ox Alpha's endpoint vanished from OpenRouter's catalogue the day the real one appeared. If you prototype on one, wire a fallback and assume the ID will stop resolving without warning.

And there is a data-handling trade that does not go away. OpenRouter's stealth terms stated prompts and completions were retained by the anonymous provider. If you sent proprietary code to Ox Alpha during the preview believing it was a routine free tier, that data sat with an operator you could not name at the time. The rule for the next stealth model, whoever runs it: open-source repositories, scratch projects and synthetic data only.

What should you do now?

If you were running on stealth/ox-alpha, the migration is a one-line model-string change to z-ai/glm-5.3-flash or glm-5.3-flash, and it is worth doing today rather than after a 404. Set reasoning_effort explicitly, turn on prompt caching if your workload resends context, and re-benchmark on your own tasks now that structured outputs are available on the production endpoint.

If you were watching from the sidelines, the interesting fact is not that a stealth model topped OpenRouter for a week. It is that the model behind it shipped with MIT weights at $0.15 per million input tokens. That combination — near-Opus agentic coding, permissive licence, Flash pricing — is what changes procurement decisions, and it is available whether or not you were in the preview. Our open-source LLM landscape guide maps where that sits against the rest of the field, and the GLM-5.3 launch guide covers the text-only sibling it was mistaken for.

One decision rule worth keeping: when a stealth model appears, evaluate it on your own tasks during the free window and write down the result — but plan your budget against the most plausible paid successor, not against $0. Ox Alpha's window was six days.

FAQ

What was Ox Alpha?

Ox Alpha was the anonymous preview name for GLM-5.3-Flash, a 320B-parameter Mixture-of-Experts model from Z.ai that activates 18B parameters per token. It ran on OpenRouter as stealth/ox-alpha from 20 to 26 August 2026 at zero cost, with a 1M-token context window and text, image and video input, before Z.ai claimed it publicly.

Who made Ox Alpha?

Z.ai, formerly Zhipu AI, the Chinese lab behind the GLM model family. It confirmed authorship on 26 August 2026. OpenRouter's archived Ox Alpha page now states directly that the model "was developed and operated by ZAI, revealed to be ZAI GLM-5.3-Flash," and Z.ai's own Ollama listing says it tested GLM-5.3-Flash anonymously as ox-alpha before release.

Is Ox Alpha still free?

No. The $0 stealth preview ended when Z.ai revealed the model on 26 August 2026. GLM-5.3-Flash now costs $0.15 per million input tokens and $0.50 per million output on list pricing, with a 50% launch promotion ($0.075 / $0.25) running until 9 September 2026. There is no :free variant on OpenRouter.

Where did Ox Alpha go?

It became z-ai/glm-5.3-flash. A check of OpenRouter's public models API on 31 August 2026 returned 396 models and no entries in the stealth/ namespace, so the old ID is no longer callable. The descriptive page at openrouter.ai/stealth/ox-alpha still loads as an archive and links to the live listing.

Is GLM-5.3-Flash open source?

The weights are openly available under the MIT licence at zai-org/GLM-5.3-Flash on Hugging Face, published 26 August 2026, with an FP8 default and a separate BF16 repository. MIT imposes no commercial-use restrictions, unlike the bespoke licences some open-weight releases carry. Training data and the full training recipe were not released.

How do I use GLM-5.3-Flash now?

Three hosted routes and one local one. Use model ID z-ai/glm-5.3-flash on OpenRouter, glm-5.3-flash against Z.ai's OpenAI-compatible endpoint at https://api.z.ai/api/paas/v4/, or glm-5.3-flash:cloud on Ollama's hosted infrastructure. For self-hosting, download the MIT-licensed weights and serve with vLLM, SGLang, KTransformers or Unsloth.

Can I run GLM-5.3-Flash locally with Ollama?

Not on your own hardware through Ollama. Ollama's library carries a single tag, glm-5.3-flash:cloud, which routes to Ollama's hosted servers in the US and Europe. To run the model on your own machine you need the Hugging Face weights plus vLLM, SGLang or a community GGUF, and enough memory for 320B parameters — roughly 160 GB at 4-bit before KV cache.

Was Ox Alpha really better than GPT-5.6?

No. The viral "80% on DeepSWE" figure came from one developer's 10-task subset, not the full benchmark. Z.ai's own published DeepSWE v1.1 score for GLM-5.3-Flash is 63.4. On Terminal-Bench 2.1 it scores 84.3 against GPT-5.6 Terra's 87.4 and Claude Opus 4.8's 85.0 — close to the frontier, not ahead of it, at roughly a ninth of GLM-5.3's price.

Why do labs release models as stealth previews?

Unbranded evaluation, production-scale load testing, real developer prompt telemetry, and a launch narrative that costs nothing. The pattern has precedent: Quasar Alpha and Optimus Alpha were OpenAI's GPT-4.1 in April 2025, and Sonoma Sky and Dusk Alpha were xAI's Grok 4 Fast in September 2025. Expect no SLA and assume the endpoint disappears without notice.