GLM-5.2: Z.ai's open-weights model, its price, its benchmarks and what replaced it

Every few months an open-weights model from China resets the conversation. DeepSeek did it, Qwen did it, and in June 2026 Zhipu AI, the company behind the Z.ai brand, did it with GLM-5.2: a Mixture-of-Experts model released under the MIT license, with a one-million-token context, that the independent benchmarker Artificial Analysis ranked first among open-weights models and fourth overall when it came out.

Four months later, the picture has moved. Z.ai shipped GLM-5.3 in August, built on the same base model, and Artificial Analysis now marks GLM-5.2 as deprecated. This article was first published on 20 June 2026 and revised on 17 October 2026: it keeps what GLM-5.2 is and what it costs, corrects what has aged, and says when it still makes sense to use it. Every figure carries its source and its date.

What GLM-5.2 is, in one table

GLM-5.2 is a sparse Mixture-of-Experts (MoE) model: most of its parameters sit idle on any given token, which is how a model this large stays affordable to run.

Spec	GLM-5.2	Source
Release	Paying users from 13 June 2026, open weights on 16 June 2026	Let's Data Science; Artificial Analysis
Parameters	753B in the published weights, about 40B active per token	Hugging Face model card; Let's Data Science (18 June 2026) for the active count
Context window	1 million tokens	Model card
Architecture	MoE with sparse attention; "IndexShare" reuses one indexer across each group of four sparse attention layers, which Z.ai says cuts per-token compute by 2.9 times at 1M context	Model card
Modality	Text (English and Chinese listed)	Model card
License	MIT, with no regional limits	Model card
Serving	SGLang, vLLM, Transformers, KTransformers, Ascend NPUs, quantized builds for llama.cpp, Ollama and LM Studio	Model card

The benchmarks: first among open models in June, superseded since

The most credible signal at launch was independent. In June 2026, Artificial Analysis placed GLM-5.2 first among open-weights models and fourth overall, behind three closed models: Claude Fable 5, Claude Opus 4.8 and GPT-5.5 (Let's Data Science, June 2026).

That ranking has not held, for two reasons. Artificial Analysis has since moved to a new version of its index, and Z.ai has shipped a successor. On 17 October 2026, Artificial Analysis gives GLM-5.2 (reasoning) a score of 34 on its Intelligence Index v4.3.2, notes that the model is deprecated and recommends GLM-5.3, which it scores at 45 on the same index.

On individual tests, keep the difference between what the vendor reports and what third parties measure. These are Z.ai's own numbers, from the model card:

Benchmark (reported by Z.ai)	GLM-5.2	GLM-5.1	Claude Opus 4.8	GPT-5.5
SWE-bench Pro	62.1	58.4	69.2	58.6
Terminal Bench 2.1 (Terminus-2)	81.0	63.5	85	84
GPQA-Diamond	91.2	86.2	93.6	93.6
Humanity's Last Exam (text)	40.5	31	49.8 (full set)	41.4 (full set)
Humanity's Last Exam, with tools	54.7	52.3	57.9 (full set)	52.2 (full set)
Source: GLM-5.2 model card, read on 17 October 2026. Vendor-reported scores; for HLE, Opus 4.8 and GPT-5.5 were measured on the full set, GLM on the text-only subset.

The honest read: on its own numbers GLM-5.2 beat GPT-5.5 on SWE-bench Pro and stayed behind Opus 4.8 on every row. For a buying decision, the independent aggregate matters more than any single vendor-reported score.

The price, and the catch

The Z.ai API price list (read on 17 October 2026) charges for GLM-5.2:

GLM-5.2 on the Z.ai API	Per million tokens
Input	1.40 dollars
Cached input	0.26 dollars (storage free "for a limited time")
Output	4.40 dollars

At launch, VentureBeat put the cost at about one-sixth of GPT-5.5 on its coding tests, a figure relayed by Let's Data Science. The ratio depends on the workload mix, so treat it as an order of magnitude.

The catch is token consumption. GLM-5.2 is a heavy reasoner: running the Artificial Analysis Intelligence Index took it 180 million output tokens, and Artificial Analysis computes a cost of 1.47 dollars per index task (page read on 17 October 2026). Cheap per token does not automatically mean cheap per job: budget for the output a model actually produces, not only for its rate card.

One more point that has changed since June: GLM-5.3 is listed at exactly the same price, 1.40 and 4.40 dollars, on the same price list. On the API, there is no saving left in staying on GLM-5.2.

Open weights mean sovereignty, not only savings

The license matters more than the price. GLM-5.2 ships under a standard MIT license, with no regional limits on the weights according to its model card. You can download the checkpoints from Hugging Face, run them on your own hardware, fine-tune them and ship them commercially. That is the difference between renting a model and owning your stack.

Self-hosting is real but not trivial. With 753 billion parameters, the FP8 weights alone weigh about 750 GB (one byte per parameter), so serving the model means a multi-GPU server node, or a quantized build with the quality trade-offs that come with it. Most teams will start on an API and keep self-hosting for the cases where it pays: strict data sovereignty, high and predictable volume, or fine-tuning on proprietary data.

This is also where GLM-5.2 keeps an edge over its successor. GLM-5.3 ships under its own "GLM-5.3 License": above 10 billion dollars of annual revenue, a security review by Z.ai is required, according to Gigazine (29 August 2026), and Artificial Analysis describes it as allowing commercial use with restrictions. For most companies that changes nothing; for a large group, plain MIT is simpler.

The catch: governance, trust and the Entity List

On 16 January 2025, the US Bureau of Industry and Security added Beijing Zhipu Huazhang Technology and nine related Zhipu entities to the Entity List, stating that they "advance the People's Republic of China's military modernization through the development and integration of advanced artificial intelligence research" (Federal Register 2025-00704). That does not stop anyone from downloading MIT-licensed weights, but it is a real signal for any organisation weighing vendor risk, and a hard constraint for some US-linked supply chains.

For day-to-day use, the convenient hosted API runs through a company subject to Chinese law. For a European business handling client or personal data, that is a governance question to answer before sending sensitive prompts. The clean answer is the one the MIT license enables: self-host the weights inside your own infrastructure, or use a provider in a jurisdiction you accept, and keep the hosted API for non-sensitive work.

The GLM line, in dates

• 16 January 2025: Zhipu entities added to the US Entity List.
• GLM-5.1: the predecessor, the comparison column in the GLM-5.2 model card.
• 13 and 16 June 2026: GLM-5.2, first to paying users, then as open weights under MIT, with a 1M context.
• 14 August 2026: GLM-5.3 launches on Z.ai's subscriptions and API (Gigazine); its model card says it uses the same base model as GLM-5.2, the gains coming from post-training (Terminal Bench 2.1: 88.2 against 81.0, vendor-reported).
• 29 August 2026: GLM-5.3 weights published on Hugging Face and ModelScope, under the GLM-5.3 License.

Our take: when to use GLM-5.2 today

What follows is our analysis.

• On the API, move to GLM-5.3. Same base model, same price, better scores on Z.ai's own benchmarks and on Artificial Analysis, which has deprecated GLM-5.2.
• Keep GLM-5.2 where the MIT license matters. If you self-host, fine-tune and redistribute, or if your company is large enough for the GLM-5.3 License condition to apply, GLM-5.2 remains the most permissive version of this model family.
• Keep sensitive data off the hosted API. Self-host, or route through a provider you have vetted. This is a governance line, not a quality one.
• Never single-vendor the model under your product. In four months this model went from first open model to deprecated. Put every model behind your own interface, so that switching is a configuration change. That is how we build our own products.

For more on how fast the ground moves under AI tools, see SpaceX buying Cursor and the suspension and return of Claude Fable 5, one of the three closed models ranked ahead of GLM-5.2 in June.