GLM-5.3-Flash Is the Cheap Open Model That Codes Like Opus 4.8
Quick verdict
Z.ai launched GLM-5.3-Flash and confirmed it was the model people had spent weeks poking at under the name Ox Alpha. It is open weights under an MIT license, 320B total parameters with only 18B active, a 1M-token context, and Z.ai says it runs entirely on Chinese chips. The number that got everyone's attention was the price. Artificial Analysis clocked it at $0.09 per task, roughly seven times cheaper than the full GLM-5.3, while scoring 57 on its Intelligence Index. Z.ai's own coding benchmark puts it level with Claude Opus 4.8. The win here is cost per token, not raw smarts, and the model is noticeably weaker on general factual knowledge than the price tag suggests. But for coding and agent work, an open model this cheap changes the math.
What actually shipped
GLM-5.3-Flash is a sparse mixture-of-experts model. The headline spec from Z.ai's launch tweet is 320B total parameters, 18B active per token, a 1M-token context, native multimodality, and an MIT license. The weights went up on Hugging Face, with API, chat, a coding plan, and AutoClaw access live the same day. Early on, Z.ai engineer Zixuan Li told people who grabbed the model on day zero to re-download it after a chat-template fix, so the first few hours had the usual launch-day roughness.
The interesting part is what got cut from GLM-5.2. Sebastian Raschka walked through the architecture and pointed out that the backbone dropped from a 744B-A40B design to 320B-A18B. It leans on a Kimi Linear-style 3:1 hybrid attention, 34 Kimi Delta Attention layers, 11 latent and sparse attention layers, a DeepSeek V4-style residual path, and a native vision encoder. His read was that both major attention components are already efficient variants, not the usual mix of one efficient and one full-attention layer. That is the lever behind serving a 1M context at a price this low: fewer active parameters, a smaller KV cache per layer, and attention cost that stops compounding as the context grows.
The price is the story
Most model launches lead with a benchmark. This one led with a receipt. Artificial Analysis scored GLM-5.3-Flash at 57 on its Intelligence Index, three points behind the full GLM-5.3 at 60, but at $0.09 per task versus $0.68 for GLM-5.3 max. That is about 7.5 times cheaper for a three-point drop. API pricing came in at $0.15 per million input tokens and $0.50 per million output, with cached input around $0.03. It ties GPT-5.6 Terra and Muse Spark 1.2 on the index while costing a fraction of either per task.
There is a catch buried in the token math, though. Artificial Analysis noted the model burned 149M output tokens to run the index, and about 90% of those were reasoning tokens. So the economics look great because the per-token price is rock bottom, not because the model is especially frugal with tokens. If you are paying by the token on a long agent run, that reasoning overhead adds up. The sticker price is real; just do not assume the model thinks in fewer words than its rivals.
Where it is strong and where it is not
The split in the reactions was clean. On coding and agent tasks, the numbers held up. Artificial Analysis put GLM-5.3-Flash at 84.3% on Terminal-Bench v2.1, just ahead of the full GLM-5.3, and gave it a GDPval agentic Elo of 1770, within margin of error of GLM-5.3 and Grok 4.6 and behind only the top Claude Opus 5 settings. Cline said it was the fastest-growing model in the tool's history, driving 11% of all traffic in under a week, and made it free in the editor. That is a real demand signal, not just a launch-day spike.
Knowledge is the soft spot. The model's AA-Omniscience accuracy was 28% with a matching 28% hallucination rate, against 47% accuracy for GPT-5.6 Terra. So it is much better at doing work in a repo than at recalling broad facts about the world. And the multimodal claim drew the sharpest pushback. skalskip92 tested it on aerial imagery, crops, technical drawings, and object detection and found it weak, arguing that "native vision" in the spec sheet does not mean the model is good at specialized vision tasks. Worth remembering before you wire it into anything that has to actually see.
The Chinese-chip angle
The line that got the most airtime was not about the model at all. Z.ai said GLM-5.3-Flash runs "entirely on Chinese AI chips," and SemiAnalysis zeroed in on the claim that 100 trillion tokens a day are being served on that hardware. teortaxesTex ran the back-of-envelope math: if each chip does around 864M tokens a day, 100T a day implies something like 116,000 chips. Nobody outside Z.ai can confirm the number, but engineers read it less as marketing and more as a statement that domestic accelerators can now carry high-volume frontier inference. That is the part with geopolitical weight, and it is why the launch landed harder than a three-point benchmark gap would suggest.
Video: GLM-5.3-Flash tested
A hands-on look at the model behind the Ox Alpha mystery, and how it holds up once you run real tasks through it.
What it means if you pay for AI
The pattern from the last GLM release repeats here. An open model that is close on coding and dramatically cheaper does not have to beat the closed frontier to matter. It just has to be good enough that renting it through a provider becomes the obvious call for a big chunk of your workload. For coding agents in particular, where you burn tokens all day, a model at $0.15 per million input tokens with agentic scores near the top tier is hard to ignore. You keep Opus or GPT-5.6 for the genuinely hard reasoning and route the bulk somewhere this cheap. That kind of model-by-task split is exactly the case for running everything through one aggregator instead of paying for a stack of separate subscriptions.
FAQ
Can I run GLM-5.3-Flash on my own machine?
Technically the weights are open under MIT, but at 320B total parameters even a heavy quant is far past a normal workstation. For almost everyone "open" means renting it from a provider like Baseten or CoreWeave, both of which announced support at launch, rather than self-hosting. See our rundown of the best AI coding agents for where models like this fit.
Is it really as good as Claude Opus 4.8?
On Z.ai's own coding benchmark, yes; on independent evals it is close on coding and agent tasks but weaker on general knowledge, with a 28% hallucination rate. Treat the Opus-4.8 parity claim as a first-party number and lean on the third-party coding scores.
Why is it so much cheaper than GLM-5.3?
It uses far fewer active parameters (18B versus the full model's larger backbone) and a hybrid attention design that keeps the KV cache small, so serving costs drop. That is how you get roughly 7.5 times lower cost per task for a three-point Intelligence Index gap.
Sources
- @Zai_org - GLM-5.3-Flash launch announcement and specs
- @Zai_org - Code Bench claim of parity with Claude Opus 4.8
- Artificial Analysis - Intelligence Index 57 and $0.09 cost per task
- Sebastian Raschka - architecture breakdown
- Cline - 11% of traffic in under a week
- SemiAnalysis - 100T tokens/day on Chinese chips
- skalskip92 - pushback on the vision claims
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix