For six days, the most used AI model on OpenRouter was one nobody would put their name on. It showed up on August 20 as "Ox Alpha": no company, no price, free for anyone, a 1 million token context window, image and video input, and suspiciously good at code. Half a million people piled in, me included, and the internet spent the week playing detective. Today Z.ai pulled off the mask: Ox Alpha is GLM-5.3-Flash, it's MIT licensed, and the full weights are sitting on Hugging Face right now.
I care because I watched it eat the charts in real time, and because the reveal finally answers the question that actually matters: what does this thing cost now that it isn't free?
The week everyone got free tokens
By the time the mask came off, the numbers were silly. Ox Alpha chewed through roughly 11.6 trillion tokens in its first 72 hours, about two and a half times the biggest launch OpenRouter had ever seen. It knocked DeepSeek off the top spot on OpenCode after a 56 day run. By reveal day the counter stood at roughly 44 trillion tokens and half a million users. All of it free, which is exactly why everyone showed up.
The detective story was the fun part. Within two days people were running fingerprinting probes against it: tokenizer quirks, an API error code specific to GLM-5.3, the exact reasoning-effort settings the GLM family uses. Every probe pointed at Z.ai, the Beijing lab behind the GLM models, while the wilder theories (Google, Microsoft, even SpaceX compute somehow) kept the timeline entertaining. So when Z.ai confirmed it to Bloomberg this morning, the mood was less shock and more "called it". The nerds had the right suspect within 48 hours.
What it actually is
GLM-5.3-Flash is a 320 billion parameter mixture-of-experts model that only activates 18 billion per token, which is how it stays cheap to run. It takes text, images and video, holds a 1 million token context window, and the whole thing, about 328 GB of weights, is MIT licensed on Hugging Face. Commercial use included, no strings. Z.ai also says the entire stealth preview ran on Chinese-made chips, which is its own headline if it holds up.
Then the price sheet: $0.15 per million input tokens, $0.50 per million output, and half that until September 9 with the launch discount. For comparison, Claude Opus 4.8 charges $5 and $25. That's roughly 30 times cheaper on input and 50 times on output, at list price, for a model that plays in a similar coding league. Even Gemini 3.7 Flash, Google's own cheap tier, costs five to seven times more.
Is it any good?
Honest answer: very good for the price, and not the new best model. Your feed will tell you it "destroys Claude and OpenAI". It doesn't. The viral 80% DeepSWE score came from one developer's ten-task run, and he walked it back himself after running the full set. Here's Z.ai's own comparison chart, so vendor numbers, bold marking the winner:
| Benchmark | GLM-5.3-Flash | Claude Opus 4.8 | GPT-5.6 Terra | Gemini 3.7 Flash |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 84.3 | 85.0 | 87.4 | 85.8 |
| DeepSWE v1.1 | 63.4 | 58.0 | 69.6 | 65.3 |
| AutomationBench | 48.8 | 41.0 | 37.2 | 52.3 |
Read that honestly and the shape is clear: it trades blows with Claude Opus 4.8 and mostly trails GPT-5.6 Terra and Gemini 3.7 Flash. The one independent number we have is Artificial Analysis's Intelligence Index, where it scored 57, level with Terra and a point above Gemini 3.7 Flash, and they ran the tests themselves. Z.ai even says openly that it isn't competing with the very top tier, Claude Fable 5 or GPT-5.6 Sol. So no, nobody got dethroned today. What changed is what this level of model costs.
The free-sample stand
The playbook here is the supermarket free-sample stand. Put the product out with no label, give it away, and count how many people come back before you tell them the brand and the price. OpenRouter has hosted these stands before: GPT-4.1 previewed as "Quasar Alpha", GPT-5 as "Horizon". The difference this time is scale. Z.ai handed out 44 trillion tokens' worth of samples, topped the charts anonymously, and only then flipped the sign around to show that the product costs a fraction of everything next to it on the shelf. As launches go, it's hard to argue with. (The store itself changed owners last week too, but that's the previous post.)
What this means in practice
If you write code: add it to your routing and test it on your own repos while the discount lasts. At $0.075 per million input tokens, running your own eval costs almost nothing, and your own eval beats every chart above. Keep a frontier model for the hard problems, use this for the volume.
If you run a team: the interesting line item is anything you run at volume: summarization, extraction, classification, internal chat. Those jobs at a thirtieth of the price is a real budget conversation. The counterweight is governance: the hosted API routes through Chinese infrastructure and Z.ai sits on a US export blacklist, so anything sensitive either stays with a Western-hosted provider or goes on the self-hosted weights, which are MIT and clean.
If you're not technical: the takeaway is that the price of "good enough AI" just dropped hard, again. Features that were too expensive to run last quarter, document Q&A over everything, per-user assistants, transcript cleanup, quietly become affordable. Expect more of them in the products you already use.
The honest caveats
Every number in that table is Z.ai grading its own homework, hours old and unreproduced; the only independent score is that 57. The 44 trillion token headline is inflated by caching, since OpenCode says about 93% of input tokens came from cache. A free week tells you what people take when it's free, not what they'll pay for. The discount ends September 9. And self-hosting "only" needs about 328 GB of memory, which is one very serious server rather than your gaming PC, though quantized builds already run on a 256 GB Mac Studio.
The mask-pull was fun, but the number I keep coming back to isn't a benchmark. It's that a model trading blows with Claude Opus now costs less per million tokens than a coffee. The top of the market didn't move today. The floor did.
Numbers from Z.ai's launch materials and Hugging Face model card, the OpenRouter listing, Bloomberg's reveal report, OpenCode's public dashboard, and Artificial Analysis's independent index run. A snapshot as of August 26, 2026, hours after the announcement: every benchmark except the Artificial Analysis score is vendor-reported, so keep the salt close until independent runs land.