GLM-5.2 Now Runs 40% of Developer Tokens on OpenRouter. Open Weights Are Not the Same as Sovereignty.

For most of the last two years the story of open weights was a story of catching up. The best downloadable model was always a generation behind the best model you had to rent. That gap is the thing that let the US frontier labs charge what they charge. This month it closed. Z.ai's GLM-5.2 now sits fourth overall and first among all open-weight models on the Artificial Analysis Intelligence Index, and on OpenRouter it is reportedly moving something like 40% of all developer tokens. The best open model on the market was trained in Beijing, on Huawei chips, and it costs a sixth of the frontier.

I want to be precise about what happened, because the headline everyone is running ("China caught up") is both true and the least interesting part. The interesting part is that a permissive license and a Hugging Face download button are being sold as sovereignty, and for the overwhelming majority of the people running this model, that is not what they are getting.

What GLM-5.2 Actually Is

GLM-5.2 is the flagship from Z.ai, the lab formerly branded Zhipu. It rolled out to paying coding customers on June 13, and then Z.ai did the thing US labs stopped doing: it published the weights under the MIT license. Not a research-only license, not a custom "open but" license with an acceptable-use annex. MIT. You can download it, run it commercially, and pay nobody a per-seat or per-token license fee to do it.

The measured numbers are what make this more than a press release. On the Artificial Analysis Intelligence Index v4.1, GLM-5.2 scores 51. That is fourth overall and first among open models, ahead of MiniMax-M3 at 44, DeepSeek V4 Pro at 44, and Kimi K2.6 at 43. On SWE-Bench Pro, Z.ai reports 62.1, which lands above GPT-5.5 at 58.6 and its own predecessor GLM-5.1 at 58.4. The context window is one million tokens. Treat the SWE-Bench figure as vendor-reported until a neutral harness confirms it, the same caveat we put on everyone. The Intelligence Index placement is not vendor-reported, and that is the number that should make a US pricing committee uncomfortable.

Trained on Huawei, Not Nvidia

Here is the part that matters for anyone who thinks export controls are holding a line. GLM-5.2 was trained on roughly 100,000 Huawei Ascend 910B processors using Huawei's MindSpore framework, with no Nvidia silicon at any stage. The Ascend 910C sits at around 60% of an H100's inference performance per a December Council on Foreign Relations report, and the training run reportedly needed about 15% more compute time than a comparable Nvidia-based run. That gap got erased by cheaper domestic chips and government subsidies. Emad Mostaque pegged the all-in training cost near $25 million, with roughly 80% of it in post-training.

Sit with those numbers. A top-four model in the world, built entirely outside the Nvidia stack, for the price of a Series A. The compute-moat thesis (that you cannot make a frontier model without tens of thousands of the best Western accelerators) is the thesis the whole export-control regime is built on, and GLM-5.2 is a live counterexample to it.

Against Opus 4.8, GLM-5.2 is roughly 82% cheaper on output. It is close to GPT-5.6 Luna on the sticker, and it is one of the few models near the top of the index whose weights you can actually hold in your hand. That combination is why it is eating token share. You can model your own workload on our cost calculator and see the model itself in the models tracker.

The 40% Number, and Why It Is a Trap

The line getting passed around this week is that GLM-5.2 now handles about 40% of developer tokens on OpenRouter while costing far less than the closed frontier. Take the exact percentage with a grain of salt, since it moves week to week and depends on how you count, but the direction is real and it is steep. A model that did not exist in public a month ago is now a plurality of a major router's traffic.

And that is exactly where the sovereignty story falls apart. "Open weights" and "I control where my data goes" are two different claims, and people are collapsing them into one. GLM-5.2 at full precision needs roughly 1.5 terabytes of GPU memory to self-host, which is on the order of nineteen Nvidia H100 80GB cards at a bare minimum. That is a data-center capital line, not a laptop and not a single server. The number of teams actually running their own GLM-5.2 cluster is small.

Everyone else routes through someone else's inference. Some of that is Western hosts like Fireworks, DeepInfra, Featherless, or SiliconFlow. But a large share, including any call that goes through Z.ai's own cloud or its ZCode product even with bring-your-own-key enabled, is processed under the jurisdiction of China's National Intelligence Law. For a workload you cannot self-host, that is a structural condition, not a hypothetical. The weights are free. The place your prompt and your codebase get processed is not something the license controls.

What I Would Actually Do With This

None of the above is a reason to avoid GLM-5.2. It is a reason to be honest about which of the three rows above you are in. The model is genuinely good and genuinely cheap, and for a large class of work that is the whole decision. If you are generating marketing copy, transforming public data, or writing code against a repo that is already open source, the jurisdiction question is close to noise, and you should be routing default traffic to it and pocketing the 82%.

The moment it stops being noise is the moment your prompt carries something you would not email to a stranger: customer records, unreleased source, anything regulated. At that point the only two rows that hold up are self-host or a Western host you have a contract with, and one of those costs real money to stand up. Do not let a MIT license badge talk you out of asking where the GPU is.

There is also a slower, larger implication for the labs charging frontier prices. The premium tier has been justified by two moats: capability and trust. GLM-5.2 just put a serious dent in the capability moat from outside the Nvidia stack and at a sixth of the price. What the frontier labs have left to sell, and what they should be selling hard, is the trust moat: a model of comparable capability whose inference runs somewhere you are allowed to send your data. That is a real product. It is just a much narrower one than "we have the smartest model," because for a growing slice of tasks, they no longer do.

Three Things I Am Watching

First, independent replication of the SWE-Bench Pro number on a neutral harness. Vendor-reported coding scores have been wrong in both directions this year, and a 62 that survives a clean rerun is a different story than one that does not.

Second, whether Western inference hosts keep serving GLM-5.2 at scale, or whether procurement and policy pressure pushes it off the big US-facing routers. If it stays, the token share keeps climbing. If it gets quietly delisted, the self-host wall becomes the whole ballgame.

Third, the US response on price. The last month has been one long argument about the pricing floor, and GLM-5.2 just set a new one from a direction export controls were supposed to prevent. Watch what the next Gemini and the next Sonnet do on the sticker, because the competitor they are now pricing against is free to download and runs on chips they cannot embargo.

Open weights are a gift. They are not a jurisdiction. Anyone selling you the second thing while handing you the first is counting on you not reading the deployment diagram. Read it.