Zhipu GLM-5.2

• 1. overview (909 chars)

  On June 17, 2026, Zhipu officially launched and open sourced the new generation flagship large model GLM-5.2. Its positioning has shifted from instant question and answer to "long-range tasks" - allowing AI to work continuously for several hours like a human and autonomously complete a complete large-scale project. On the Artificial Analysis comprehensive list, GLM-5.2 scored 51 points, ranking in the top three together with Anthropic and OpenAI, forming the so-called "New Royal Three" pattern. This is the first time that an open source model competes with closed source flagships in the most hard-core engineering scenarios such as code and long-range tasks. The model adopts a 744B parameter MoE architecture (approximately 40B active parameters per token). It is completely open source under the MIT protocol and can be used for free for commercial use. Neither training nor inference relies on overseas computing power. It has been adapted to nine major domestic computing platforms including Huawei Ascend, Pingtou Ge, and Moore Threads on the first day of launch.

• 2. Product background (1386 chars)

  Zhipu (Beijing Zhipu Huazhang Technology Co., Ltd.) is the leading large-scale model startup company in China. It disclosed its first performance report on the Hong Kong Stock Exchange in March 2026 (stock code 02513.HK), and has entered the commercialization stage from the pure research and development stage. GLM-5.2 is another major iteration of GLM series on GLM. Starting from the beginning of 2025, Zhipu has invested almost all its efforts in researching coding capabilities. It has gone through continuous iterations of GLM-4.5, GLM-4.7, GLM-5.0, and GLM-5.1, and finally achieved the goal of "open source model coding SOTA" with GLM-5.2. The core breakthrough of the model is to make millions of context windows available for engineering use. GLM-5.2 proposes the IndexShare architectural innovation—reusing the same indexer between every four sparse attention layers, reducing the FLOPs per token to 2.9 times under the context length of 1M. The MTP layer for speculative decoding has also been improved, increasing the accepted length by up to 20%. The training side relies on the self-developed Slime framework to support large-scale Agentic RL and OPD training. In terms of industry positioning, the release of GLM-5.2 coincides with the critical time window when the strongest overseas models are turning to closed and the demand for open source alternatives is increasing. Claude Fable 5 is subject to export controls by the United States, and overseas developers are in urgent need of reliable high-end alternatives. The emergence of GLM-5.2 has filled the market gap, and is called "DeepSeek moment 2.0" by some analysts.

• 3. Product functions and usage experience (2480 chars)

  The core capabilities of GLM-5.2 revolve around three major directions: Long Horizon Task, Coding, and Ultimate Infra Optimization. For long-range missions, GLM-5.2 implements the Solid 1M lossless context. This is not only marked with 1M on the parameters, but also can be used in engineering in actual applications. Zhipu spent several months to significantly expand the training environment of 1M Coding Agent, covering many typical fields such as large-scale implementation, automation research, and performance optimization. On the FrontierSWE test set (a benchmark that examines whether AI can complete complex technical projects on a scale of hours to tens of hours), GLM-5.2 scored 74.4%, only 1% lower than Claude Opus 4.8 and exceeding GPT-5.5's 72.6%. According to a Zhipu demonstration, GLM-5.2 can process more than 880,000 tokens in a round of continuous tasks, independently complete the complete software delivery process from development, joint debugging, testing to packaging and online, and produce a complete application covering web pages, mobile terminals and mini programs within a few hours. In terms of coding capabilities, GLM-5.2 remains open source SOTA on multiple authoritative benchmarks. In the Design Arena human preference evaluation, GLM-5.2 ranked first in the world with 1360 Elo, slightly surpassing Claude Opus 4.8's 1350 Elo. It scores 62.1% on SWE-bench Pro, exceeding GPT-5.5’s 58.6%. It scored 81.0 on Terminal-Bench 2.1 (a benchmark that evaluates AI Agents operating computers through the command line), an improvement of 17.5 percentage points compared to GLM-5.1. On MCP-Atlas (a tool usage review), GLM-5.2 is only 0.8% different from Opus 4.8. The model also introduces effort level (thinking gear) control, so users can balance capabilities, speed, and cost. In terms of Infra optimization, GLM-5.2's online inference relies on multiple domestic computing platforms. On Day 0, it has completed inference adaptation with domestic computing platforms such as Huawei Ascend, Pingtou Ge, Moore Thread, Cambrian, Kunlun Core, Muxi, Haiguang, Biren, Tianshu Zhixin, etc., achieving stable operation with high throughput, low latency, and large concurrency on domestic chip clusters. However, in actual use, user feedback also revealed several obvious shortcomings. Some developers have tested it and pointed out thatIt turns out that GLM-5.2 has a "dispersion" problem - it is easy to forget the context and go off track in multi-step tasks. There are also users who report that the model has a tendency of "illusory expansion requirements". The user asks for A, and the model automatically adds B and C, increasing the review burden. In addition, GLM-5.2 is a text-only model and does not have native multi-modal capabilities - the 2026 flagship model does not have vision, which looks awkward in the context of competing products generally having multi-modal capabilities.

• 4. Pricing and business model (1196 chars)

  API pricing for GLM-5.2 is extremely competitive. On a per million token basis, the input cost is approximately $1.40 and the output cost is approximately $4.40. For comparison, the input/output costs of Claude Opus 4.8 are $5/$25 respectively, and GPT-5.5 is $5/$30. Calculated based on a simulation scenario of 50 million tokens/month, the monthly cost of GLM-5.2 is about $145, Opus 4.8 is about $750, and GPT-5.5 is about $875. Choosing GLM-5.2 can save up to $730 per month. At the same time, GLM-5.2 is completely open source under the MIT protocol, and enterprises can build and deploy it themselves, completely eliminating the cost of API billing by token. This is particularly attractive to companies with strict data compliance - the MIT protocol can be freely modified for commercial use, domestic deployment data does not leave the country, and there is no fear of interruption of external service supply. However, GLM-5.2 has a "high consumption" problem under Zhipu's own package system. Due to the high-end positioning of the model, the quota is consumed at a multiple rate when calling (3 times during peak periods, 1 times during non-peak periods). In addition, the single inference token consumption is large (all 1M contexts are pulled in). Ordinary Lite package users reported that "the 5-hour quota is used up in one and a half hours", and the actual usage experience is limited by the package restrictions.

• 5. User reviews (1314 chars)

  In a sample of 807 real users, 69% gave positive reviews. Positive feedback is concentrated in three directions: outstanding front-end development capabilities, superior cost performance to similar closed-source models, and high flexibility of the MIT open source protocol. A developer migrated the entire Opus workflow to GLM-5.2, and the computing power cost dropped from US$186 to US$17. An enterprise blindly tested the effect of front-end landing page generation and found that the quality of the finished product was almost the same as that of Opus 4.8, but the cost was only 1/6. There are even Fortune 500 companies that have moved half of their coding efforts to this model. Judging from user portraits, 82% of users are programmers, focusing on technical scenarios such as coding, agent development, and local deployment, and 71% are English users. This proves that the popularity of GLM-5.2 is purely a global developer circle, rather than pan-C-end traffic. The negative feedback was equally clear. About 31% of real users gave varying degrees of negative reviews. The main complaints include: slow inference speed, low token efficiency (the same task consumes more tokens than competing products), insufficient running stability (occasional errors such as "The prompt parameter was not received normally"), and lack of visual ability. A Nuggets blogger concluded after using it for a day that "it's OK to write code, but don't expect it to plan complex events." He also shared his combination plan of "Claude as the brain and GLM-5.2 as the hand."

• 6. Industry evaluation (1476 chars)

  Reviews from industry media were generally positive. Mainstream media such as Science and Technology Daily, CCTV, and China Daily have made extensive reports on GLM-5.2, emphasizing the significance of its combination of "open source national model + domestic computing power". A report in China Daily pointed out that "In the context of the strongest overseas models turning to closed and the demand for open source alternatives rising, this combination has attracted attention from the entire industry." Tang Jie, the founder of GLM, confirmed in an internal letter "The Big Wave Is Coming" on July 11 that multiple core indicators of GLM-5.2 have equaled or even exceeded closed-source flagships such as GPT-5.5. This announcement confirms an industry transition that is taking place - for the first time, open source models are competing with closed source flagships in the most hard-core engineering scenarios such as code and long-term tasks. In terms of competitive product landscape, direct competitors of GLM-5.2 include Claude Opus 4.8 (Anthropic), GPT-5.5 (OpenAI), Kimi K2.7 Code (Dark Side of the Moon), MiniMax M3, etc. On the coding track, Claude Opus 4.8 still maintains the lead overall, but GLM-5.2, as an open source model, has reduced the gap with the closed source flagship to within 5%. In terms of comprehensive cost performance, GLM-5.2 has an overwhelming advantage. However, some analysts pointed out that the "fire" of GLM-5.2 has a certain false high component. A full analysis of 9,163 related tweets shows that only 9% of users have actually used it and given specific reviews, and nearly 80% are driven by the volume of communication and capital effects. The popularity of the topic is much higher than the actual user size.

• 7. Controversies and Risks (1320 chars)

  The main disputes and risks faced by GLM-5.2 focus on several aspects. The first is the practical stability issue of long contexts. Although the official claim is that "Solid 1M is lossless context," many developers have found in actual tests that the model suffers from distraction during multi-step long-distance tasks - "It doesn't matter if the window is stretched to 1 million, the brain still only looks at the beginning and the end." GLM-5.2 is not immune to the collective shortcomings of long context models. Secondly, there are issues of product stability and user experience. Some users reported that when using the official Coding Plan of Zhipu, even reading a txt file would result in an error. The error message was `reason=unknown`, and there was a lack of troubleshootable error prompts. There is a gap between the capabilities of the model itself and the degree of completion at the product level. The third is the controversy over billing policy. The doubling mechanism of 3 times consumption during peak periods has caused many users to complain that the quota "burns too fast", and ordinary packages can hardly support real development scenarios. This issue triggered extensive discussions in the community that "the model is strong but it is fun to use it, which is two different things." Finally, there is the loss of visual ability. The 2026 flagship model does not have native multi-modal capabilities. In the context of competing products generally having multi-modal capabilities, this shortcoming may limit the application scenarios of GLM-5.2.

• 8. Applicable groups and usage suggestions (959 chars)

  GLM-5.2 is most suitable for programmers and development teams, especially in the following scenarios: high-frequency coding automation requirements; large code warehouses that require millions of contexts; teams that are cost-sensitive and want to reduce API fees; companies that have data compliance requirements and require local deployment; developers who want to migrate their coding work from a closed-source model to an open-source model. The least suitable scenarios include: tasks that require multi-modal understanding (picture viewing, video analysis, etc.); complex and fuzzy high-level architecture designs that require the model to independently plan; and interactive scenarios that require extremely high real-time reasoning speed. For ordinary developers, the recommended usage is to "leverage strengths and avoid weaknesses": hand over refactoring and coding tasks that require a global vision and clear goals to GLM-5.2, and hand over links that require intention understanding, requirement clarification, and complex planning to models that are good at "thinking clearly" such as Claude or GPT.

• 9. Summarize (675 chars)

  GLM-5.2 is a landmark breakthrough for domestic large models in the field of high-end coding. For the first time, it puts the open source model on the same track as the closed source flagship in terms of code and long-range tasks, and provides developers with a truly implementable alternative with extreme cost performance and the MIT open source agreement. At a turning point in the industry where the strongest overseas models are turning closed and the domestic computing power ecosystem is maturing, GLM-5.2 is on the right track. But it is not impeccable - the attention stability of long context, the lack of multi-modal capabilities, the completion of the product level, and the user experience issues caused by the billing strategy are all the "next mountains" that it needs to continue to conquer.

No in-depth report yet

• 

  SLeeK

  —

  用了一周 GLM-5.2，最大的感受就是编程能力确实强，前端生成页面基本跟 Opus 一个水平，但确实有涣散的问题，长任务跑到后面就忘了前面说过什么。

---

• 

  Adrianc41

  —

  从 5.1 换到 5.2，最明显的感觉就是上下文真的大了，以前 20 万 token 撑满就幻觉，现在整个项目塞进去还能改得动。

---

• 

  Stephanie_Morgan_Pro

  —

  当天额度一个小时就烧完了，我还以为是 bug，结果发现是 5.2 在高峰期按 3 倍扣额度，太难了。

---

• 

  DAOthinker464

  —

  对比 Opus 4.8 和 GPT-5.5 实测了一周，前端能力 GLM-5.2 确实不虚，但复杂系统的架构设计还是 Claude 更强，5.2 容易在中间步骤跑偏。

---

• 

  CAphi

  —

  把整套 Opus 的工作流迁过去了，算力成本从 186 降到 17 美元，真的很香。

---

• 

  reddog874

  —

  开源协议是 MIT，可以随便商用部署，这点对合规要求高的公司真是太友好了。

---

• 

  BenjaminFlores_Pro

  —

  纯文本模型，2026 年了旗舰机不带视觉，说实话有点尴尬，同期的 Kimi K2.7 Code 和 MiniMax M3 都是多模态。

---

• 

  r6xxu1c

  —

  试了一下让它读 txt 文件，直接报错「The prompt parameter was not received normally」，太离谱了。

---

• 

  Sam_anthaWood

  —

  写代码确实不错，但让它做规划就别指望了，我的用法是 Claude 当大脑、GLM-5.2 当手，配合起来体验最佳。

---

• 

  z5l3kae7k

  —

  价格确实是降维打击，按 5000 万 token 算一个月才 145 美元，Opus 要 750，这差价足够让我忍它的缺点了。

---

• 

  段娜

  —

  1M 上下文是真的能用的，不像某些模型标了百万其实几十万就开始掉链子，5.2 我塞了八十几万 token 跑完一整个工程没断。

---

• 

  SPeterson_2020

  —

  让我最烦的是它老爱自作主张，让它做 A 它非要顺手把 B 和 C 也做了，审起来比自己做还累。

---

• 

  CarolJ_ohnson

  —

  全球可用模型 Code Arena 第一，这成绩确实硬，不是吹的。

---

• 

  co25ko0ac3

  —

  公司里盲测了前端落地页生成，成品跟 Opus 4.8 几乎没有肉眼可见的差距，但成本只有六分之一，果断切了。

---

• 

  Aaron.Cook_77

  —

  Lite 套餐根本扛不住 5.2，一个半小时见底，升了 Pro 才好点。

---

• 

  JOols

  —

  推理速度偏慢，等得有点心焦，但考虑到这个价格也认了。

---

• 

  MHoward_2021

  —

  在 Claude Code 里接上它，改了一个几万行代码的项目，十轮交互下来不跑偏，体验比 5.1 好太多了。

---

• 

  GameFiGamer273

  —

  Vercel 的 CEO 都说「almost shocked」，这波国产模型是真的支棱起来了。

---

• 

  WPowell_2024269

  —

  Fable 5 被 ban 那天 GLM-5.2 刚好开源，时间点卡得太妙了，MIT 协议意味着谁也 ban 不了它。

---

• 

  CatherineHolm

  —

  用 Rust 从零复刻阿波罗登月计算机那个 demo 真的震撼到我了，虽然日常用不上，但足以证明它的实力。

---

• 

  Nicholas_Murphy_77

  —

  SWE-bench Pro 62.1 超了 GPT-5.5 的 58.6，这个数据挺说明问题，但跟 Opus 4.8 的 69.2 还有差距，营销话术说「接近 Opus」多少有点夸张了。

---

• 

  Emily_Henderson_66

  —

  高峰期 14 点到 18 点千万别用，3 倍扣额度真扛不住，我现在都赶上午干活。

---

• 

  James.Sanchez_2022

  —

  国产算力适配是亮点，华为昇腾、摩尔线程都能跑，不用担心被卡脖子。

---

• 

  CrnptoLink

  —

  前端开发能力确实强，Design Arena 排第一不是吹的，生成页面审美在线，但后端那种需要理解复杂业务逻辑的活它就不太行了。

---

• 

  Michellew70

  —

  token 消耗太大，同个任务比 Claude 多用 50% 的 token，虽然单价便宜但总价优势没那么大。

---

• 

  TheStephanieAnderson_dev

  —

  知乎大 V 说「以后用 Opus 可能其实是 GLM-5.2 冒充的你都分不出来」，笑死。

---

• 

  Kenneth_MendozaJr10

  —

  试了让它写一个 Minecraft 克隆，直接跑出来了能飞的版本，体验比 GPT-5.5 的版本还好。

---

• 

  sadmeercat181

  —

  自己部署的话门槛不低，744B 的大模型需要挺多显存，不是人人都能本地跑的。

---

• 

  SUpet

  —

  个人开发者小团队用性价比无敌，闭源模型那种按 token 付费的模式对高频调用太伤了。

---

• 

  Bruce_Kelly

  —

  24 小时跑完一个 SaaS 项目从开发到上线，88 万 token，实现了我对 AI 编程的终极想象。

No reviews yet