Ox Alpha turns out to be GLM-5.3-Flash

Here's the announcement in numbers. GLM-5.3-Flash has 320 billion (320B) total parameters, but only 18 billion (18B) of them switch on to generate each token. On Artificial Analysis's independent composite intelligence index, it scores 57 at a cost of $0.045 per task. The weights were released on Hugging Face under an MIT license, which allows unrestricted modification and commercial use, and Z.ai says that for the past week, every bit of this model's traffic was served not on NVIDIA GPUs but on Chinese-made AI chips. If some of these terms sound unfamiliar, don't worry — we'll unpack them one by one below.

It hid its name and took the top spot first

The most dramatic part of this whole story is how the model was rolled out. Starting August 20, Z.ai quietly listed this model for free under the name "Ox Alpha" on the model marketplace OpenRouter and the coding tool OpenCode, without disclosing who made it, where it came from, or how big it was. Developers were left to judge it purely on performance, with no idea about the company behind it.

The result was overwhelming. According to OpenCode's own tally, Ox Alpha processed 43 trillion tokens in a single week and took the top usage spot. The runner-up, a DeepSeek-family model (deepseek-v4-flash), handled 22 trillion tokens, while the model ranked tenth, a GPT-5.6-family model (gpt-5.6-luna), handled 827 billion — meaning Ox Alpha beat second place by nearly double. During that stretch, it served 497,000 unique users across 12.85 million completed coding sessions.

The guessing game over its identity got heated too. On Reddit and X, tokenizer analysis (examining how the model splits text into pieces) fueled speculation that it belonged to the GLM family, while others argued it was an unreleased Microsoft model. Stripe CEO Patrick Collison, whose company is in the process of acquiring OpenRouter, added fuel to the fire by calling it "very impressive". METAL LAB reported on August 23 that our token-breakdown analysis matched GLM-5.2 on all 60 counts — and that guess turned out to be right. Bloomberg's report confirmed Z.ai as the maker, and the official blog post followed a few hours later.

In other words, the model was validated first through real-world usage data, stripped of any brand or nationality bias, and only revealed its identity after it had already earned a place inside developers' everyday tools.

Only 18B active out of 320B — the specs made simple

Parameters are the numbers an AI model stores from training to hold onto what it's learned. Generally, more parameters mean a smarter model, but also a costlier one to run. GLM-5.3-Flash gets around this using an architecture called "mixture of experts" (MoE). It splits its 320 billion parameters into groups of specialists, and every time it generates a word, it only switches on the roughly 18 billion parameters belonging to the specialists relevant to that particular task. Think of a hospital with 3,200 doctors on staff, but any given patient only ever sees the 18 specialists they actually need. The design keeps the total knowledge base large while keeping the computation used at any one moment small.

This is also the first native multimodal model in the GLM-5 series. Rather than only reading text, it can handle images and video within the same model. That opens up workflows like building a webpage and then visually checking the rendered result to fix it directly. Its context window — how much it can hold and work with at once — is 1 million tokens, roughly enough to load about ten full-length novels at a time and ask questions about them.

The license is MIT, one of the most permissive open-source licenses available: anyone can download the weights, modify them, plug them into their own product, and make money off it. Z.ai says training used roughly 30 trillion tokens of multimodal data.

Performance is closing in on Claude Opus 4.8

The benchmark scores Z.ai published make the model's standing pretty clear. A benchmark, for anyone unfamiliar, is basically a standardized exam that gives every model the same test to score them against each other.

Benchmark	What it measures	GLM-5.3-Flash	GLM-5.2 (predecessor)	Claude Opus 4.8	GPT-5.6 Terra
Terminal Bench 2.1	Completing server setup and debugging tasks end-to-end in a terminal	84.3	81.0	85.0	87.4
DeepSWE v1.1	Real-world coding evaluation using tasks from private repositories	63.4	46.2	58.0	69.6
AutomationBench v1.0.6	Performing office automation tasks	48.8	26.2	41.0	37.2
GDPval-AA v2	Real work tasks across 44 professions (evaluated by Artificial Analysis)	1773	1504	1582	1571

DeepSWE tests models on tasks from repositories that have never been made public, which lowers the odds that a model could have memorized the answers during training (data contamination). Here, GLM-5.3-Flash scored 63.4, edging out Claude Opus 4.8's 58.0. On GDPval-AA, which measures real professional work, it scored highest among all the models compared, and its office-automation score nearly doubled its predecessor's. That said, the gap with the very top models is still visible — on both Terminal Bench and DeepSWE, it falls short of GPT-5.6 Terra.

There's an interesting detail buried in Z.ai's own coding evaluation. On Z.ai's internal Code Bench, run on top of Claude Code, GLM-5.3-Flash — when set to its maximum "thinking effort" — scored 29.0, nearly matching Claude Opus 4.8's 29.5. But getting there took nearly 140,000 output tokens per task, compared with about 120,000 for Opus 4.8. Meanwhile, the higher-tier Claude Fable 5 scored a much higher 39.5 using around 110,000 tokens. In short, the setup lets a cheaper model think longer to approach the answers of a pricier one. Since the per-token price is so low, the total cost still comes out far cheaper even with all that extra "thinking."

Worth noting: most of these figures come from Z.ai's own testing. Still, independent numbers point the same direction. On Artificial Analysis's composite intelligence index — which rolls coding, math, and reasoning into a single score — GLM-5.3-Flash scored 57. According to Z.ai, that score bracket has historically belonged to models costing roughly 10 times as much per task. That's exactly what the chart in the hero image shows: Alibaba's Qwen3.8 Max sits at the same 57-point level but on the expensive right edge of the chart, while GLM-5.3-Flash sits alone in the cheap bottom-left corner.

Pricing — half off until September 9

The API price is $0.15 per million input tokens and $0.50 per million output tokens. A token is the unit models use to count text — for Korean text, roughly one to two tokens per character is a decent rule of thumb. Cached input, which gets reused when the same prompt repeats, is priced at $0.03. According to Z.ai's official pricing page, there's currently a launch promotion cutting all of that in half — $0.075 input, $0.25 output, $0.015 cached input — running until midnight Singapore time on September 9.

The comparison puts things in perspective. The higher-tier model GLM-5.3, launched two weeks earlier, prices its API at $1.40 input and $4.40 output — meaning GLM-5.3-Flash is roughly a tenth of that even at full list price. Subscribers to Z.ai's GLM Coding Plan get three times the usage limit compared with GLM-5.3.

How it got this cheap — inside the architecture

The secret behind this pricing lies in the architecture. Compared with Z.ai's own GLM-4.5 (355B, a similar total parameter count), the number of active parameters dropped from 32B to 18B, and the number of layers (the stacked computation stages inside a model) was cut roughly in half, from 92 to 45. Instead of stacking deep, the model goes wide and shallow, cutting how much gets switched on each time.

The core trick is cutting attention costs. Attention is the computation a model uses to compare every word against every other word to understand relationships in the text — and that cost balloons exponentially as the text gets longer. For a document a million tokens long, that cost becomes unmanageable. GLM-5.3-Flash solves this by blending two kinds of attention. Linear attention compresses everything that's already been read into a fixed-size summary it carries forward, while sparse attention uses an index to pull up only the specific parts it needs right now. Instead of rereading the entire book every time, it's like carrying a summary notebook around and only flipping to the relevant page in the index when it actually needs that detail. On top of that, it adds a compression technique called IndexPool, which squeezes four of the index's keys down into one.

The payoff: at a million-token context length, attention computation dropped 3.0x and the KV cache shrank 4.4x compared with the flagship GLM-5.3. The KV cache is temporary memory the model stores during a conversation so it doesn't have to recompute earlier parts from scratch — and at long context lengths, this is the main thing that eats up GPU memory. Shrinking both numbers means the same hardware can serve more users at once, and that difference is exactly what shows up in the price.

The real story: it ran entirely on Chinese chips

Z.ai says every bit of traffic that flooded in through Ox Alpha over the past week was served on clusters of Chinese-made AI chips. The company built its own inference engine on top of the open-source inference framework SGLang, running it across tens of thousands of domestic accelerators, and split the workload — image/video encoding, prefill (reading the prompt), and decoding (generating the answer one token at a time) — across separate server pools handling each stage. It's a division-of-labor setup, similar to a kitchen where prep, cooking, and plating are each handled by a different team.

Chinese chips generally trail NVIDIA GPUs in memory capacity and bandwidth, which becomes a real problem when handling a million-token context. Z.ai says it compensated with quantization — storing weights and cache at 8 bits or lower to shrink their footprint — and with layer-splitting, which spreads the model across multiple chips. The company claims this tripled end-to-end throughput on the same hardware compared with its early setup, and pushed the per-token cost down to a level comparable to mainstream NVIDIA GPUs. Worth flagging: that cost claim is currently just the company's own statement, not independently verified.

There's another detail worth paying attention to here: a GLM-5.3-based infrastructure agent reportedly helped with this optimization work. That means the model itself was writing kernels (low-level code that controls the chips), hunting down bottlenecks, and tuning the very system running it. There's long been a saying that NVIDIA's real moat isn't the chips themselves but a decade-plus of accumulated CUDA software — and this suggests China is now throwing models, not just engineers, at closing that gap. It becomes a feedback loop: as the model improves, the domestic chip software improves faster, which in turn makes running the model even cheaper.

NVIDIA reports earnings the same day

From a market perspective, this announcement touches two nerves. The first is pricing. Once there's an option to get 57-point intelligence at $0.045 per task, there's less reason to hand repetitive workloads to frontier models charging $1–2 per task. Splitting hard problems off to closed frontier models while routing routine work to cheap open models is already becoming standard practice at companies — a trend that lines up exactly with the recent shift we reported, where developers started splitting work over to GLM after Claude Fable's price hike.

The second is chips. When DeepSeek rattled the market in January 2025, it wasn't really about performance — it was the revelation that training costs were far lower than assumed. This time, the claim is about inference, the side of AI that involves actually running the model day to day, being pulled off on Chinese-made chips. If training cost pressured model pricing before, replacing inference chips is a story that could shift where NVIDIA's demand actually comes from.

Coincidentally, NVIDIA reports its fiscal 2027 second-quarter earnings after the U.S. market close the same day. Company guidance points to roughly $91 billion in revenue, and that forecast assumes zero revenue from China-bound data center compute. So on the same day, NVIDIA has to prove it can keep growing without the Chinese market, while a Chinese model maker claims it just pulled off frontier-level inference without NVIDIA at all — two opposing answers landing on the same day.

Of course, this still needs to be proven out. Absorbing 43 trillion tokens during a free promotion is one thing; sustaining that same load at the same speed after switching to paid usage is another. Whether throughput and response speed hold up after the free promotion ends on September 9 will be the first real test of this claim.

Editor's take

What flips the script here is the order of the reveal. If Z.ai had opened with "this is a Chinese model," plenty of developers probably wouldn't have bothered trying it — instead, an unnamed model pulled in 43 trillion tokens over a week. Letting performance speak first and revealing nationality only afterward turned into something like an experiment showing just how thin preconceptions about Chinese AI really are once real usage gets involved.

But the line that will probably outlast the spec sheet is this one: "served entirely on Chinese chips, at a per-token cost on par with NVIDIA." It's still just a company's own claim for now, but if it holds up, this isn't really a story about model pricing anymore — it's a story about the shape of global compute. It's a sign that the feedback loop of models closing their own software gap has already started spinning. That's reason enough to keep an eye on the throughput numbers once the free tier ends on September 9.