Every few weeks a new "cheap but capable" model shows up and asks for your attention. Most don't earn it. GLM-5.2, released by China's Zhipu AI (Z.ai) in mid-June 2026, is one of the ones that does — because it isn't trying to win a leaderboard. It's trying to be the model you can actually afford to leave running: a coding-first, tool-calling workhorse with a genuine 1-million-token context window, open weights under an MIT license, and a bill that stays quiet even when your agents don't.

If you just want to call it, you have two doors. You can sign up directly with Z.ai, or reach it through a model aggregator such as OrcaRouter, which exposes GLM 5.2 through a standard OpenAI-compatible endpoint. Either way, the rest of this guide is what you'll want to know first: what it is, what it costs, how it scores, and where it actually fits.

TL;DR — GLM-5.2 is Z.ai's flagship for long-horizon, agentic coding. 1M context / 128K output, text-only, hybrid reasoning, native tool calling, MIT open weights. $1.40 in / $4.40 out per million tokens (cache reads $0.26). Not the #1 frontier model — but for the price, one of the best value-per-token options shipping today.

GLM-5.2 at a glance

• Developer: Zhipu AI / Z.ai — model id z-ai/glm-5.2

• Released: mid-June 2026, on the GLM Coding Plan, with standalone API and MIT-licensed open weights

• Context window: 1,000,000 tokens; up to 128K output in a single call

• Type: text-in / text-out, hybrid reasoning via reasoning_effort (high / max), native tool calling & JSON

• Price: $1.40 / 1M input, $4.40 / 1M output, cache reads $0.26 / 1M

• Built for: repo-scale agentic coding, multi-step engineering workflows, long-horizon delivery

Why GLM-5.2 is interesting

The headline is the context window. A million tokens isn't a spec-sheet flex here — it's roughly a 700,000-word working memory. You can drop an entire codebase, a 500-page compliance manual, or a hundreds-of-turn support transcript into a single prompt and let the model reason over all of it without the chunk-summarize-restitch dance that quietly eats your engineering time.

The second thing is that GLM-5.2 was built coding-first. It launched on Z.ai's GLM Coding Plan and is tuned for the messy, multi-step reality of real software work: read the repo, plan the change, call the tools, run the task to completion. Deep reasoning is on by default, and you dial the effort up with a single parameter when a problem is genuinely hard.

And then there's the part that makes procurement smile: open weights under an MIT license. If compliance or data-residency rules ever force you off a hosted API, you can self-host the same model instead of rewriting your stack around a different one. Cloud when it's convenient, on-prem when it's required — same weights, same behavior.

Benchmarks: how good is it, really?

Let's be honest about where GLM-5.2 lands: it isn't chasing the very top of the overall leaderboards. What it does instead is punch far above its price. On Artificial Analysis it scores 56.8 on the Coding index (#19 of 106 models) and 60.8 on the Intelligence index (#14 of 110) — better than roughly 8 in 10 models tracked, at a fraction of frontier pricing.

Task-level numbers back that up: SWE-bench Pro 62.1, Terminal-Bench 2.1 up to 82.7, GPQA Diamond 74.4, and a near-perfect AIME 2026 of 99.2 on math. (All figures are vendor/aggregator-reported — as always, run your own eval on your own tasks before you bet a product on a leaderboard.)

The short version: if your workload is high-volume coding, extraction, or long-document work, GLM-5.2 gives you frontier-adjacent quality without frontier economics.

GLM-5.2 pricing

This is where it gets fun. At $1.40 per million input tokens and $4.40 per million output, GLM-5.2 costs a fraction of the Western flagships — Claude Opus 4.8 is $5 / $25, GPT-5.5 is $5 / $30. Cache reads drop to $0.26 per million, and through a zero-markup gateway you pay Z.ai's exact rate.

To put it in real money: a workload of 10 million tokens a month lands around $23 — roughly a large latte, dropping to about $19 with prompt caching. A single request with 100K input and 5K output costs about $0.16. This is the kind of pricing where you stop rationing calls and just let the agent work.

How to access the GLM-5.2 API

If your code already speaks the OpenAI format, switching to GLM-5.2 is a two-line change, not a rewrite. Point your base URL at the provider, set the model string, and go:

That's the whole migration. Because the endpoint is OpenAI-compatible, the same call works from the OpenAI SDK, LangChain, LlamaIndex, or tools like Cursor and Cline — and calling through a gateway means you can A/B GLM-5.2 against Opus, DeepSeek or Gemini from one dashboard and one key. Full specs and a live playground are on the GLM-5.2 model page.

Who should use GLM-5.2 — and who shouldn't

Reach for it when: you're running coding agents, processing whole documents or repositories, or doing high-volume extraction and summarization where the monthly bill — not the single hardest prompt — is what keeps you up at night. The 1M window and low price make it a natural default for that work.

Look elsewhere when: you need the single best score on the very hardest reasoning problems, or your prompts comfortably fit in 32K-128K tokens and you're cost-sensitive — a smaller, cheaper model will be faster and lighter for that. GLM-5.2's edge is long-context, coding-heavy, high-volume work; use it where that edge is real.

FAQ

Suggested reading: