GLM 5.2 Just Launched: 1M Context, Coding-First, Open Weights Next Week (Day-One Brief)

What is GLM 5.2?

GLM 5.2 is the latest member of Zhipu AI / Z.ai's flagship Chinese-developed model family, a direct successor to GLM-5.1 (which shipped on a $18/mo Coding Plan baseline earlier this spring). Z.ai is rolling 5.2 out as a coding-first model — every public claim around the launch is about agentic coding tasks, long-horizon refactors, and using the full million-token context for repository-scale work.

The release happened today, June 13, 2026, and is unusual in two ways:

• Z.ai shipped access first, paperwork second: the model is live across all Coding Plan tiers (Lite, Pro, Max, Team) immediately, while 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. There is no SWE-bench Verified score, no LiveCodeBench result, no HumanEval. Vendor marketing positions 5.2 as superior to 5.1 on coding and long-horizon tasks, but third-party verification is pending.

What does Zhipu confirm about GLM 5.2?

Context window

The headline number is 1,000,000 tokens (1M). The full-window model ID is glm-5.2[1m]. Maximum output is capped at 131,072 tokens — wide enough for full pull-request-scale diffs and long agentic plan-then-execute traces.

Thinking effort

Z.ai exposes two thinking-effort presets, High and Max. Zhipu's own guidance is that Max should be the default for coding work. There is no "Auto" or "Low" tier — both presets aim to be slow-and-thoughtful by default, which fits the long-horizon framing.

Coding Plan pricing

GLM 5.2 is available on every tier of the existing GLM Coding Plan, with the same prompt-weekly caps as 5.1:

• Lite — around 400 prompts / week, ~$18 / month baseline (this is the floor and where most individual devs land).
• Pro — around 2,000 prompts / week.
• Max — around 8,000 prompts / week.
• Team — seat-based organisation pricing.

If you already subscribe to a tier, you have GLM 5.2 now at no extra cost.

Agent compatibility

Z.ai shipped 5.2 with first-day support across the major OSS agentic-coding CLIs and IDE wrappers:

• Claude Code (Anthropic's official CLI)
• Cline (formerly Claude Dev)
• OpenCode
• Roo Code
• Goose
• Crush
• OpenClaw
• Kilo Code

If you are already driving one of these agents off another model, swapping in GLM 5.2 is a config change rather than a workflow change.

What is NOT confirmed about GLM 5.2?

Day-one launches always leave gaps. The honest list of what is not public at this moment:

• No benchmarks. No SWE-bench Verified, no LiveCodeBench, no HumanEval, no AIDER polyglot score. Independent third-party numbers are not yet out either.
• No parameter count. GLM-5 was a 744B-parameter mixture-of-experts. The 5.2 architecture is not specified in the launch materials — could be the same backbone with additional training, could be different.
• No open weights yet. They are promised under the MIT license but had not appeared on Hugging Face as of the announcement. The repo to watch is the zai-org/glm-5 family.
• No standalone API yet. If you are not on the Coding Plan you cannot try the model directly today — the API is part of next week's drop.
• No technical report. No data-mix details, no fine-tuning recipe, no evaluation methodology.

We will update this post as each of those drops next week. The shape of the launch — early access on the paid plan, weights later, no benchmarks at launch — is consistent with how Z.ai shipped 5.1 in March, so the timing is plausible.

How does GLM 5.2 fit next to GLM-5.1 and the broader 2026 landscape?

Helpful context for understanding the upgrade:

GLM-5 (February 2026)

The original GLM-5 launched February 11, 2026 with 744B MoE parameters. Independent SWE-bench Verified coverage put it at 77.8% — competitive with frontier closed models at the time.

GLM-5.1 (Spring 2026)

5.1 introduced the Coding Plan, a self-reported coding score at ~94.6% of Claude Opus 4.6's number, and the open-weights MIT release that drew most of the community attention. See our GLM-5.1 local-run guide for the practical setup.

GLM 5.2 (today)

5.2's framing is incremental on capability — same Coding Plan, same agent ecosystem, same tier prices — but the new 1M context is the upgrade that is most likely to matter in practice for repo-scale agentic coding. "Strong long-horizon" is also Z.ai's headline language; we will know what that means concretely when the benchmarks land.

For broader landscape framing — how Kimi K2.7, GPT-5.5, Claude Opus 4.8, DeepSeek V4 stack up against each other in mid-2026 — start with our Kimi K2.7 vs GPT-5.5 vs Claude Opus 4.8 comparison. We will fold GLM 5.2 numbers into that grid as soon as third-party evaluations show up.

Who is GLM 5.2 actually for today?

Based on what is shipped right now, the realistic answer is two groups:

1. Existing GLM Coding Plan subscribers. If you are already paying ~$18/month for Lite, you have a brand-new model with a 1M context window in your pocket as of today. Try it on a multi-file refactor your other agent stumbled on.
2. Teams shopping the open-weights model market. The MIT-licensed weights next week make 5.2 a candidate for self-hosting (the same way GLM-5.1 became a popular on-prem coding model). If you are still on GLM-5.1 weights, hold off on a migration until the 5.2 weights and at least one independent benchmark land.

Anyone outside those two buckets — devs without the Coding Plan subscription, teams committed to closed-model APIs (Anthropic, OpenAI), or anyone who needs verified benchmarks before adopting a model — should bookmark this post and check back next week.

How do I try GLM 5.2 right now?

The only way today (June 13, 2026) is through the GLM Coding Plan. The flow is:

1. Sign up at z.ai and subscribe to any Coding Plan tier (Lite at ~$18/month gets you in).
2. Configure your agent of choice (Claude Code, Cline, OpenCode, etc.) to point at the GLM endpoint your tier provides.
3. Use the model ID glm-5.2[1m] for the 1M-context variant. Default the thinking effort to Max for coding tasks.

Once next week's drop happens you will also be able to (a) call a standalone API without the Coding Plan, (b) chat in the Z.ai web app, and (c) pull MIT-licensed weights for local or on-prem inference. We will update this post the moment any of those land.

FAQ

When did GLM 5.2 release?

June 13, 2026 — today. The release happened across Z.ai's existing GLM Coding Plan; the standalone API, the Z.ai chatbot, and the open-source weights are scheduled for next week.

Is GLM 5.2 open source?

Not yet. Zhipu announced MIT-licensed open weights are coming next week. As of launch (June 13, 2026) the model is only accessible through the Coding Plan.

What is GLM 5.2's context window?

1,000,000 tokens with model ID glm-5.2[1m]. Maximum output tokens cap at 131,072 — wide enough for repo-scale agentic refactors and long plan-then-execute traces.

What are the GLM 5.2 benchmarks?

Zhipu did not publish any benchmark numbers at launch. No SWE-bench Verified, no LiveCodeBench, no HumanEval, no AIDER polyglot. Independent third-party benchmarks are pending. We will fold them into this post as they appear.

How much does GLM 5.2 cost?

The Coding Plan baseline is ~$18 / month (Lite tier, ~400 prompts/week). Pro is ~2,000 prompts/week, Max is ~8,000 prompts/week, Team is seat-based pricing. GLM 5.2 is included on every tier at no extra cost.

How does GLM 5.2 compare to Claude Opus 4.8 or GPT-5.5?

Until Zhipu or an independent group publishes benchmark numbers, the honest answer is "unknown." For 2026 context: GLM-5.1 self-reported ~94.6% of Claude Opus 4.6's coding score (never independently verified). Read our Kimi K2.7 vs GPT-5.5 vs Claude Opus 4.8 comparison for the broader frontier-model grid.

Which AI coding agents work with GLM 5.2?

Day-one support: Claude Code, Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, Kilo Code. If your agent already speaks an OpenAI-shaped chat-completions API and supports custom endpoints, GLM 5.2 should drop in as a config swap.

Can I run GLM 5.2 locally?

Not yet. The MIT-licensed weights are promised for next week. Until then the only way to use the model is through Z.ai's Coding Plan. When weights ship you can run them locally the same way you would run GLM-5.1 locally on CPU/GPU.

What was GLM-5.1?

GLM-5.1 is the previous flagship in this line. It introduced the Coding Plan structure and shipped open weights under the MIT license, which is how 5.2 is structured too. Read our GLM-5.1 local-run guide for the practical setup.

What about Kimi K2.7?

Kimi K2.7 from Moonshot AI is the other major open-weights flagship in mid-2026. We have a complete guide to Kimi K2.7 and the K2.7 vs GPT-5.5 vs Claude Opus 4.8 comparison. Once GLM 5.2 weights drop next week, expect a head-to-head in the same series.

Update — June 15, 2026 (48 hours post-launch)

The day-one brief above stands; here is what 48 hours of community testing, vendor docs and reviewer posts have now confirmed (or quietly left pending).

Confirmed

• Coding Plan tier pricing — Lite $10/mo, Pro $30/mo, Max $80/mo, plus seat-based Team. GLM 5.2 access is included at every tier. Quarterly billing drops the same tiers to roughly $27 / $81 / $216 per quarter under the current Q2 2026 promo.
• Drop-in agent integrations shipped live at launch via the OpenAI-compatible endpoint: Claude Code, Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, Kilo Code. For Claude Code it is a three-line change in settings.json.
• Two thinking-effort levels: High and Max. No Low / Auto. Thinking mode adds roughly 30-80% to first-token latency and roughly halves throughput on long runs.
• Architecture inheritance from GLM 5/5.1: 744B total parameters / ~40B active per token, 384 experts, DeepSeek Sparse Attention powering the 1M-token window, 28.5T pretrain tokens.
• Hardware sizing for self-host when the weights drop: FP8 weights ~800 GB on disk → 8× H200 SXM (1,128 GB HBM) is the production sweet spot with KV cache headroom for the full 1M context. INT4 quants ~200 GB run on 4× H200 with a 1-3% coding-benchmark regression.

Still pending

• MIT open weights not on Hugging Face yet. Track huggingface.co/zai-org for the GLM-5.2 repo and a matching GLM-5.2-FP8 companion — that matches the GLM 5.1 release pattern.
• Standalone per-token API not yet live on open.bigmodel.cn or z.ai/pricing. GLM 5.1 standalone is the reference at $1.40 input / $4.40 output per M tokens; expect GLM 5.2 to land near or below that.
• Hosted-provider endpoints (Together, Fireworks, DeepInfra, Groq, OpenRouter) — none list GLM 5.2 yet because the weights are not public. Expect 3-10 day catch-up after the MIT drop; Fireworks and DeepInfra were first on 5.1.
• chat.z.ai chatbot still serves GLM 5.1 in the free tier. The 5.2 chatbot rollout is part of the same "next week" batch.
• Cursor, Continue, Aider integrations — not yet wired. Cursor has an open community thread requesting GLM-5 support but no merged work.

Independent benchmark status

Still none on the standard suites. As of 48 hours post-launch no third party has published SWE-bench Verified, SWE-bench Pro, LiveCodeBench, Terminal-Bench 2.0, AIDER Polyglot, GPQA Diamond, or HumanEval scores specifically for 5.2. Anyone quoting an SWE-bench number for 5.2 right now is conflating it with GLM 5.1's published 58.4 SWE-Bench Pro / 63.5 Terminal-Bench 2.0 / 71.8 MCP-Atlas line. Treat the 5.1 numbers as the credible floor for 5.2 capability until independent runs land.

Community sentiment after 48 hours

The Hacker News reception thread (269+ points, 146 comments) consolidated into two consistent reads:

• Bull case: "punches above its weight" on UI/design code and modern coding conventions; the 1M window is the upgrade most likely to matter in production (no more file chunking). One commenter described shipping a non-trivial GTK/Rust/Lua app where "GLM wrote 93%." Throughput on early hosted providers is in the 35-70 tok/s band.
• Bear case: "about six months behind the frontier labs, similar to Opus in January" on multi-file architectural reasoning. Run-to-run variance and harness sensitivity (Terminal-Bench swung 40.4% → 48.3% on GLM 5 depending on the agent wrapper) are unresolved carry-overs.

Top-comment verdict: test it today if you are already on the Coding Plan; do not rebuild your stack around it until third-party benchmarks land next week. That matches our take.

Legal / compliance notes that materially affect a switch

• MIT, when it ships, is fully permissive — no field-of-use, no MAU threshold, no AUP.
• Zhipu is on the US BIS Entity List since January 15, 2025. Using MIT open weights is not a regulated export under current EAR readings, but US federal customers and most defense primes treat Chinese-origin models as effectively blocked regardless of license.
• EU AI Act: GPAI model with likely systemic-risk-tier compute. Zhipu has not signed the GPAI Code of Practice and has not published an Annex XI model card — downstream EU deployers in regulated industries carry the full transparency burden.

Companion reads we just published

• GLM 5.2 vs Claude Opus 4.8: Should You Switch Your Coding Stack?
• GLM 5.2 vs DeepSeek V4: The Open-Weights Coding Showdown
• GLM 5.2 vs GPT-5.5: Open-Weights vs Closed Flagship for Coding
• Kimi K2.7 vs GLM 5.2: Two Chinese Open-Weights Flagships Compared