Zhipu AI’s Z.ai just dropped GLM 5.2, their latest flagship model, and it lands with a couple of big numbers and one noticeable absence.
The model is live right now across every GLM Coding Plan tier — Lite, Pro, Max, and Team. The headline feature is a usable 1-million-token context window, which is a serious step up from the 200K context on GLM-5 and GLM-5.1. Maximum output is capped at 131,072 tokens, enough for full PR-scale diffs and long agentic traces.
But here is what makes this launch interesting: Z.ai shipped access first and the paperwork second. The standalone API, the Z.ai chatbot, the MIT-licensed open weights, and the technical report are all scheduled for next week.
No benchmark numbers were published at launch either — no SWE-bench Verified, no LiveCodeBench, no HumanEval. Vendor claims of “powerful coding capabilities” and “strong long-horizon” performance are unverified for now.
What GLM 5.2 Brings
The big picture is straightforward. GLM 5.2 is a coding-first model aimed at agentic engineering. Every public claim Z.ai has made about this launch revolves around coding tasks, long-horizon refactors, and using the full million-token window for repository-scale work.
Key specs that are confirmed:
- Context window: 1,000,000 tokens input (model ID
glm-5.2[1m]) - Max output: 131,072 tokens
- Thinking effort: Two presets — High and Max. Zhipu recommends Max as the default for coding work. No Auto or Low tiers.
- Agent compatibility: Day-one support for Claude Code, Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, and Kilo Code
- Pricing: Same Coding Plan pricing as GLM-5.1. Existing subscribers get 5.2 at no additional cost
The thinking-effort design is worth noting. Both High and Max are “slow and thoughtful” by default, which lines up with the long-horizon framing. Z.ai wants this model thinking hard on every call, not racing through quick completions.
The Open-Source Bet
The MIT-licensed open weights arriving next week are the real story here. Z.ai has been consistent about open-sourcing their model family — GLM-5 and GLM-5.1 are both on Hugging Face under MIT — and GLM 5.2 continues that pattern.
This matters because the open-weight coding model space is getting crowded fast. I covered Kimi K2.7 literally yesterday, and that’s just the latest competitor in a field that already includes MiniMax M2.5, Qwen3-Coder, DeepSeek V3.2, and the rest of the GLM family.
Every one of these is pushing SWE-bench scores into the high 70s and low 80s, narrowing the gap with proprietary models.
GLM 5.2’s 1M context gives it a differentiator that most of those competitors do not have right now. Whether that translates into better real-world coding performance is something we will know once the benchmarks and weights land next week.
What Is Missing
It is worth being upfront about what we do not know yet:
- No benchmark scores of any kind
- No parameter count or architecture details (GLM-5 was a 744B MoE; 5.2 is unspecified)
- No open weights yet (repo is
zai-org/glm-5on Hugging Face, same as previous releases) - No standalone API (only accessible through the Coding Plan today)
- No technical report (data mix, fine-tuning methodology, evaluation)
That is not a knock on the model. Z.ai shipped access first and documentation next week. But it means today’s news is more of a promise than a proof. Independent benchmarks will tell the real story.
How It Fits
GLM 5.2 sits on top of a fast-improving model family. GLM-5 shipped in February 2026 as a 744B MoE with 77.8% SWE-bench Verified. GLM-5.1 followed in April with 58.4 on SWE-Bench Pro and a focus on long-horizon agentic tasks, demonstrating multi-hour optimization runs and 600-iteration coding loops. GLM 5.2 now adds the 1M context on top of that trajectory.
The framing is incremental on capability — same Coding Plan, same agent ecosystem, same tier prices — but the 1M context is the upgrade that will matter most for repo-scale agentic coding. If you have ever tried to feed an entire codebase into a model’s context window, you know the difference between 200K and 1M is the difference between a single module and the whole project.
Who Should Care
If you are already on a GLM Coding Plan, you have access right now. Go try it on a multi-file refactor and see how the 1M context changes what you can do in a single session.
If you are shopping for open-weight models, the MIT release next week makes this worth watching. But wait for independent benchmarks before migrating anything critical.
If you are following the open-source AI landscape, the pace is accelerating. Kimi K2.7 yesterday. GLM 5.2 today. The gap between open and proprietary is shrinking faster than most people realize.
I will update this post when the weights, API, and benchmarks land next week.
