GLM 5.2: Zhipu’s 1M-Context Open-Source Model Explained

Update, September 25, 2026: GLM 5.2 is now a legacy checkpoint. The GLM Coding Plan no longer serves it and routes plan requests for it to GLM 5.3. On the Artificial Analysis Intelligence Index v4.3.2 it scores 33.7, behind Z.ai’s own GLM 5.3 at 44.8, Kimi K3 at 43.6 and the MIT-licensed GLM 5.3 Flash at 41.8. The MIT weights and the standalone API rate card are unchanged, and the benchmark and pricing sections below have been rebased.

GLM 5.2 arrived on June 13, 2026 with three headline numbers. It is a 744-billion-parameter Mixture-of-Experts design with a 1-million-token context window, under an MIT license that makes the weights free for anyone to download and run. Chinese AI lab Zhipu AI (operating as Z.ai) shipped it as a coding-first frontier model, and the timing was not subtle. It went public just two days after the US ordered Anthropic to cut foreign access to its Fable 5 and Mythos 5 models. If you want the bigger picture, see how GLM measures up against other leading models like Claude, GPT and Gemini.

GLM 5.2 is no longer the GLM to reach for, and the clearest signal comes from Z.ai itself, which stopped serving it on its own subscription. It is still the largest GLM released under a plain MIT license, and the weights are still there to download. Our roundup of the best open source AI models tracks who leads the category now. If the category itself is new to you, our explainer on what open source AI is covers how open weights and model licensing actually work. Below you’ll find the specs, how it differs from GLM 5.1, the benchmarks against rivals like ByteDance’s Seed 2.1 Pro, and what the GLM Coding Plan really costs.

The Key Takeaways

• GLM 5.2 launched June 13, 2026 as a 744B-parameter Mixture-of-Experts model with 40B active parameters per token.
• The 1-million-token context window is roughly 5x larger than GLM 5.1’s ~200K limit, with output up to 131,072 tokens.
• It ships under an MIT license, with open weights, a standalone API, and the Z.ai chatbot arriving the week after launch.
• The GLM Coding Plan now meters usage in credits and starts at $18/month (Lite), with Pro at $80/month and Max at $168/month, and annual billing takes 30% off. It no longer runs GLM 5.2.
• GLM 5.2 scored 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 2.1 at launch, ahead of GPT-5.5 and behind Claude Opus 4.8. On the Intelligence Index v4.3.2 it now sits at 33.7, behind GLM 5.3, Kimi K3 and GLM 5.3 Flash.

What Is GLM 5.2?

GLM 5.2 was the flagship large language model from Zhipu AI, a Beijing-based lab that has become one of China’s most aggressive open-weight model publishers, until GLM 5.3 replaced it in August 2026. It sits in Zhipu’s full GLM model family. It is built on the GLM-5 base and uses a Mixture-of-Experts architecture with 744 billion total parameters, of which only 40 billion activate for any given token. That design keeps the running cost closer to a 40B model while drawing on the knowledge of a much larger one.

The release is squarely aimed at developers. Zhipu positioned GLM 5.2 around long-horizon coding and agentic work, the kind of multi-step tasks where a model writes, runs, and revises code across an entire project. It connects to popular coding clients through an Anthropic-compatible endpoint, so tools like Claude Code, Cline, OpenCode, and OpenClaw can point at it without a proprietary SDK.

One thing changed fast after launch. Zhipu shipped GLM 5.2 with no published benchmark scores, as noted in launch coverage from MarkTechPost, which left early “beats GPT-5” claims as vendor assertions. Within days, independent testers and the company’s technical card filled that gap, and the numbers landed well above the earlier GLM 5.1, which scored 58.4 on SWE-bench Pro. The full results are in the benchmark section below.

GLM 5.2 vs GLM 5.1: What Changed

The jump from GLM 5.1 (released April 7, 2026) to GLM 5.2 is incremental on paper but meaningful in one dimension, context. The new model handles five times more input, which changes what you can feed it in a single prompt. Here is the side-by-side.

The two reasoning modes are new and practical. High handles everyday generation, while Max is the setting Zhipu recommends for complex, multi-step coding, trading speed for deeper reasoning. The parameter count did not move between the two releases. Both are 744B total with 40B active, so per-token inference cost is unchanged. What Zhipu changed is the architecture around it. The model card names IndexShare, which reuses one indexer across every four sparse attention layers and cuts per-token FLOPs by 2.9x at a 1-million-token context. It also cites an improved MTP layer for speculative decoding that lifts acceptance length by up to 20%.

GLM 5.2 Benchmarks: How It Actually Performs

The launch left a hole, but the data filled it quickly. Zhipu’s own technical card and independent testers put GLM 5.2 close to the closed frontier on long-horizon coding. On some tests the gap to Claude Opus 4.8 narrowed to roughly a single point, as detailed in coverage from The Decoder. That framing has moved several times since. Opus 4.8 gave way to Claude Opus 5 on July 24, 2026, then to Claude Fable 5.1 on September 1, 2026, and Claude Opus 5.5 took the flagship spot on September 22, 2026. On the Artificial Analysis Intelligence Index v4.3.2, read on September 25, 2026, Opus 5.5 leads at 57.6 and Fable 5.1 sits at 53.4, while GLM 5.2 scores 33.7, behind Kimi K3 at 43.6 and behind Z.ai’s own GLM 5.3 at 44.8. The table below keeps the comparison Zhipu actually ran in June 2026, against Opus 4.8 and GPT-5.5.

The standout result is SWE-bench Pro, where GLM 5.2’s 62.1 beats GPT-5.5 at 58.6 and its own predecessor at 58.4, while trailing Claude Opus 4.8 at 69.2. On Terminal-Bench 2.1 it leapt to 81.0 from GLM 5.1’s 63.5 on the Terminus-2 harness, and on FrontierSWE it sits under a point behind Opus 4.8 while edging out GPT-5.5. Harness choice matters more than it sounds. Artificial Analysis, running Terminal-Bench 2.1 itself that summer, scored GLM 5.2 at 77.9 against 85.0 for Kimi K3 and 84.6 for Opus 4.8, so read any single figure with the harness attached. It also posts a 99.2% on the AIME 2026 math test. For context on how far orchestration can push these numbers, Sakana AI’s Fugu reaches 73.7 on SWE-Bench Pro by routing tasks across a pool of models rather than running as one.

The economics were the real headline in June 2026. GLM 5.2 matched or beat GPT-5.5 on these coding marathons for roughly one-sixth the cost, which reframed it from cheap alternative to serious contender. The one asterisk came from independent platform Artificial Analysis, which found the model burns through far more tokens than rivals to reach those scores, making it one of the less token-efficient options in its class. That asterisk now decides the whole question. Artificial Analysis puts the cost of running its index on GLM 5.2 at $0.96 per task against $0.25 for GLM 5.3 Flash, which scores eight points higher. On hard, single-shot reasoning tests like Humanity’s Last Exam it still trails Opus 4.8 and Gemini 3.1 Pro by several points, so this is a coding-and-agents specialist first.

Why the 1-Million-Token Context Window Matters

A 1-million-token context window means GLM 5.2 can hold roughly 750,000 words of input at once. In practical terms you can paste an entire mid-sized codebase, a full set of API docs, or a long research corpus and ask the model to reason over all of it without chunking. The model is labeled glm-5.2[1m] to flag the extended window, and Z.ai describes the capacity as “usable” rather than a marketing ceiling.

For coding specifically, this is the difference between feeding the model one file and feeding it the whole repository. Agentic tools that previously had to summarise or retrieve slices of a project can now keep far more in working memory. That reduces the errors which creep in when a model loses track of code it cannot see. That long-context strength is the same reason open models built for running on a Mac have grown popular with developers who want local control.

GLM 5.2 and the Export-Control Backdrop

The launch date matters. In the days just before the GLM 5.2 announcement, the US Commerce Secretary ordered Anthropic to block foreign access to Fable 5 and Mythos 5 within 48 hours under a new export-control directive. Anthropic disabled both models globally within hours, and you can read the full timeline in our coverage of the Fable 5 shutdown.

Zhipu framed GLM 5.2 as a direct counterweight. By releasing a frontier-class model under an MIT license with no regional restrictions, the company pitched open weights as insurance against any single nation or vendor controlling foundational AI. Investors responded fast, and Zhipu’s stock rose on the open-source announcement, according to the South China Morning Post. It is part of a broader wave of Chinese open releases that includes the Qwen family and DeepSeek V4.

GLM 5.2 Pricing and Availability

GLM 5.2 went live immediately across every GLM Coding Plan tier, with the standalone API, the Z.ai chatbot, and the downloadable MIT weights all released in the week after launch. That has changed twice since. The plan moved to a credits-based system on July 30, 2026, replacing the old per-prompt allowances, and it no longer serves GLM 5.2 at all: Z.ai routes plan requests for GLM 5.2 and GLM 5.1 to GLM 5.3. What is left is the same Anthropic-compatible endpoint, a rolling 5-hour window and a weekly cap, with annual billing taking 30% off each tier. Credits are deducted by token multipliers that vary by model, and off-peak usage costs half the standard rate, with peak hours running Monday to Friday, 14:00 to 18:00 Singapore time.

GLM Coding Plan Lite ($18/month)

The entry tier costs $18/month, or $151.20 a year, which works out at $12.60/month. It covers 2,000 credits every 5 hours and 10,000 a week, which suits individual developers doing moderate coding assistance. It is the cheapest way to run GLM 5.3 inside a tool like Claude Code.

GLM Coding Plan Pro ($80/month)

Pro runs $80/month, or $672 a year at $56/month, and raises the ceiling to 12,000 credits every 5 hours and 60,000 a week. It is aimed at power users running multi-file refactors and agentic tasks daily.

GLM Coding Plan Max ($168/month)

Max costs $168/month, or $1,411.20 a year at $117.60/month, and allows 28,000 credits every 5 hours and 140,000 a week, for developers who keep an agent running through the workday. Team pricing is seat-based for organisations. For a sense of how that stacks up against rival open models, our breakdown of Rio 3.5 Open covers another Chinese open-weight release worth comparing.

Should You Use GLM 5.2?

If you write code and want a long-context model you can run yourself, GLM 5.2 is still the largest GLM carrying a plain MIT license, which removes the usual restrictions on commercial use. What has changed is that it is neither the only self-hostable GLM any more nor the best one. Its successor, GLM 5.3, published its weights in August 2026, though under Z.ai’s own licence rather than MIT, and GLM 5.3 Flash is a smaller 320B multimodal model that is MIT, outscores GLM 5.2 by eight points on the Intelligence Index, and costs a quarter as much to run. Kimi K3 scores higher than GLM 5.2 too, but it ships under a custom Moonshot licence and weighs 1.56 TB to download against GLM 5.2’s 1.51 TB. The 1M-token context is still a real advantage for repo-scale work, and the Anthropic-compatible endpoint means you can slot it into existing tools with minimal friction.

The benchmarks have settled the performance question, so the real considerations now are cost shape and data jurisdiction. GLM 5.2 is token-hungry, which means heavy agentic use can run up usage faster than the headline price suggests. And while the open weights let you self-host with full privacy, routing through the hosted Z.ai API sends your code to servers governed by Chinese data rules. That is a tradeoff worth weighing for sensitive or proprietary work. If you would rather not manage models at all, apps like Fello AI give you Claude, ChatGPT, Gemini, Grok, and DeepSeek through one $9.99/month subscription. You can compare outputs without juggling separate accounts or API keys. GLM 5.2 itself joined that lineup in the Fello AI 6.7.0 update. On a Mac it runs as a native GLM desktop client, so there is no browser tab and no API key to manage.

Conclusion

GLM 5.2 was a clear statement, a frontier-class, 1M-context, MIT-licensed model shipped within days of a major US export clampdown. And unlike at launch, the numbers backed the framing. It beat GPT-5.5 on SWE-bench Pro and closed to within a point of Claude Opus 4.8 on long-horizon tasks, all for roughly a sixth of the cost. What has changed is the company it keeps: Kimi K3 and Z.ai’s own GLM 5.3 are both ahead of it now, and Claude Opus 5.5 is the closed-side flagship two generations past Opus 4.8. The remaining caveats are its heavy token appetite and the data-jurisdiction question around the hosted API. If you want GLM 5.2 specifically, the paths left are the standalone API and the open weights, because the $18/month Coding Plan Lite now answers with GLM 5.3. Zhipu’s next release, GLM 5.5, is already expected later in 2026.

FAQ