After anonymously topping the charts for a week, Zhipu GLM-5.3-Flash is unveiled, matching Opus 4.8 in capability, with a surprise on price
After a week of topping the charts anonymously, Zhipu GLM-5.3-Flash is revealed, with capabilities matching Opus 4.8 and a surprising price point. A week ago, a model quietly appeared on OpenRouter, named Ox Alpha, which the Chinese community dubbed "Niu Lai" (literally "the bull has come"). No one knew who it belonged to. But it hit the ground running without any explanation, burning through 23 trillion tokens on OpenRouter in just a few days.
💡 What You Will Learn
After a week of topping the charts anonymously, Zhipu GLM-5.3-Flash is revealed, with capabilities matching Opus 4.8 and a surprising price point. A week ago, a model quietly appeared on OpenRouter, n
📜 Table of Contents
After a Week of Anonymous Dominance, Zhipu's GLM-5.3-Flash Revealed, Matching Opus 4.8's Capabilities at a Surprising Price
A week ago, a model quietly appeared on OpenRouter called Ox Alpha, nicknamed "Niu Lai" (The Bull Has Come) in Chinese communities.
No one knew who made it. But it hit the ground running, burning through 23 trillion tokens on OpenRouter in a few days, and snatching the top spot on OpenCode from DeepSeek, which had held it for 56 consecutive days. It topped the call volume charts on both platforms.
At 9 PM on August 26th, Zhipu stepped forward to claim it. Ox Alpha is GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with weights open-sourced directly.
Actually, before the reveal, the community had already cross-referenced three factors—tokenizer characteristics, inference parameters, and API error patterns—and was 99% certain it was Zhipu. I followed this for an entire week, routing our team's daily tasks through it. The conclusion is simple: it matches Opus 4.8's capabilities on one end, and undercuts the post-price-hike DeepSeek V4 Flash on the other. Neither competitor has any room left.
From Anonymity to Reveal: What Happened This Week
On August 20th, Ox Alpha appeared anonymously on OpenRouter.
At 9 PM on August 26th, Zhipu officially launched and open-sourced GLM-5.3-Flash.
In those seven days, developers worldwide voted with their usage. On OpenCode, it processed 42 trillion tokens in six days, setting a single-day usage record for the platform at 4x the previous peak. Zhipu also revealed a more striking detail: all of Ox Alpha's traffic during that period ran on 100,000 domestic chips—Huawei, Hygon, and Moore Threads—without a single Nvidia GPU.
Three Things I Did With It
Coding. I gave a scaffolding requirement from an internal project to GLM-5.3-Flash, GLM-5.3, and Claude Opus 4.8, asking each to write a version. Flash produced a complete result in one pass, and its code structure clarity actually surpassed GLM-5.3.
Image Understanding. I threw a poster with a glass cube effect at it and asked it to replicate it using Three.js. It recognized MeshPhysicalMaterial on its own—the module in Three.js specifically for simulating real-world materials like glass and metal—and filled in all the parameters itself: refractive index 1.44, dispersion 1.0, clearcoat 1, plus it added a film grain shader and a layer of blue haze. This kind of task used to be exclusive to the most expensive closed-source models.
General Capability. On third-party evaluation platform Artificial Analysis's Intelligence Index, it scored 57, tying Opus 4.8. In the past, that score meant a dilemma: either pay Opus 4.8's price, or put up with DeepSeek V4 Flash's "cheap but capability-limited" situation after its August price hike.
GLM-5.3-Flash dismantled that dilemma.
Four Mao per Million Tokens
| Model | Input (CNY/million tokens) | Output (CNY/million tokens) |
|---|---|---|
| GLM-5.3-Flash (limited-time discount) | 0.4 | 1.4 |
| GLM-5.3-Flash (standard) | 0.8 | 2.8 |
| GLM-5.3 | 8 | 28 |
| DeepSeek V4 Flash (post-hike off-peak) | 1.5 | 4.5 |
| Claude Opus 4.8 | ~35 | ~175 |
The limited-time discount ends at midnight on September 9th, after which it returns to 0.8 and 2.8 yuan. Even at full price, it still undercuts DeepSeek V4 Flash by a tier and sits at one-tenth of GLM-5.3's standard price.
Zhipu wrote one line in its release: "Same intelligence, 1/40th the price. Frontier intelligence, for the first time, doesn't need to be rationed."
I ran the numbers against the price table—that's no exaggeration.
The Confidence Behind the Low Price Lies in Architecture
GLM-5.3-Flash has 320B total parameters, activating only 18B per inference. For comparison, GLM-4.5 has 355B total parameters with 32B activated. The activation scale is nearly halved, yet performance surpasses earlier GLM-5 series models.
That confidence comes from three architectural changes. A hybrid architecture combining sparse attention and linear attention means the model doesn't need every token to attend to every other token, cutting attention computation to one-third. Manifold-constrained hyperconnections (mHC) restructure the network's connectivity. KV cache—the memory space in VRAM that stores conversation context—is compressed to roughly 22% of its original size.
Combined with pretraining on 30T tokens of multimodal data, it natively handles text, images, and video. Visual capabilities are also woven into the coding process—when building frontends, games, or Blender 3D scenes, it proactively checks the rendered output before continuing to iterate.
Where to Use It
Three entry points are already live.
Zhipu BigModel open platform offers direct API calls billed per token. Z Code is Zhipu's self-developed coding agent, giving new users 5 days with 5 million tokens per day. GLM Coding Plan is subscription-based with Lite, Pro, and Max tiers; the new version switches to a points system, charging standard points during peak hours (2 PM to 6 PM on weekdays) and half points at all other times.
Weights are simultaneously available on Hugging Face under the MIT license, compatible with domestic chips—if you want local deployment, you can download them directly.
My Recommendation
If your main model is DeepSeek V4 Flash, you're probably already looking for an alternative after the price hike. If it's Opus 4.8, that price is hard for individual developers to use liberally. GLM-5.3-Flash is the only option this week that cuts the barrier down to four mao at the 57-point capability tier.
I've already switched my daily tasks over.
This is Uncle Peng, the Big Player. The Ox Alpha testing for this week wraps up here. Next, I'll be tracking its real call volume on OpenCode and OpenRouter—whether 100,000 domestic chips can handle long contexts will become clear in early September.
Which domestic model are you using to replace Opus 4.8 right now? Share your thoughts in the comments.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
