GLM 5.2란? 1/6 가격으로 제공되는 오픈 웨이트 코딩 모델

The one-sentence version

GLM 5.2 is Z.ai's open-weight flagship language model, released June 13, 2026, built specifically for coding, reasoning, and tool-driven "agentic" work — the kind of multi-step tasks where a model plans, calls tools, reads results, and revises across a long session.

Z.ai is the international brand of Zhipu AI, a Beijing research company that spun out of Tsinghua University's Knowledge Engineering Group in 2019. "Open-weight" is the load-bearing phrase: the model's actual parameters are published on Hugging Face (under zai-org/GLM-5.2) and on ModelScope and Ollama, under an MIT license with no regional restrictions. You can self-host it, fine-tune it, and ship it in a commercial product without asking anyone.

Why an open-weight coding model is a bigger deal than the benchmarks

Before the mechanism, the motivation. The reason this release got attention isn't that it's the smartest model in the world — it isn't. It's that it closed most of the gap to the closed frontier while being free to download and cheap to call. For a developer, that changes the math on two decisions that used to be settled.

The first is lock-in. If your coding agent runs on a closed API, you cannot run it offline, you cannot inspect it, and your pricing is whatever the vendor decides next quarter. Open weights remove all three constraints at once. The second is cost. Reported API pricing for GLM 5.2 is $1.40 per million input tokens and $4.40 per million output tokens, which Z.ai positions at roughly one-sixth the cost of comparable frontier models. For a workload that burns tokens — and agentic coding burns a lot of them — that ratio is the whole story.

The catch, and there's always a catch: the open weights are safe to self-host, but routing your data through Z.ai's cloud API means it travels through infrastructure subject to China's National Intelligence Law, and the US Department of Homeland Security has warned that framework could compel Chinese companies to hand over data on US persons. The two facts coexist — free, inspectable weights you can run anywhere, and a hosted API with a real data-jurisdiction question. Which one applies to you depends entirely on whether you self-host or call the cloud. I'll come back to this.

How it works, without the hand-waving

GLM 5.2 is a Mixture-of-Experts (MoE) model. The reported size is about 744 to 753 billion total parameters — sources disagree slightly, which is itself a sign the precise number is still settling — with only around 40 billion active for any given token.

That split is the central trick, so it's worth one analogy. A dense model is like a single generalist who has to think about everything for every question. An MoE model is more like a large firm: it holds the knowledge of a very big organization, but for any one task it only wakes up the few specialists who are relevant. You get the capacity of a 744-billion-parameter model at roughly the serving cost of a 40-billion one. Compared to its predecessor GLM 4.5 — 355B total, 32B active — GLM 5 scaled the firm up (to 744B / 40B) and trained it on more data (28.5 trillion tokens, up from 23 trillion).

Three other pieces matter, and each exists to solve a specific problem rather than to pad a feature list.

The first is a sparse-attention design Z.ai calls IndexShare. The problem it solves: attention cost grows painfully as the context window gets long, and GLM 5.2's window is very long (more on that below). Normally a model recomputes which earlier tokens to attend to at every layer. IndexShare computes that index once at the first of every four attention layers and reuses it for the next three. Z.ai reports this cuts the dot-product indexing cost by 75% in those reused layers, and per-token compute by about 2.9× at the full one-million-token context length. In plain terms: it's what makes a million-token context affordable to actually run.

The second is dual reasoning modes — two selectable thinking-effort settings called High and Max. Max is for hard, multi-step coding where the model needs room to plan and revise; it can consume close to 85,000 output tokens on a single task. High gives up only a few points of performance while roughly halving that token output, which is the lever you reach for when latency and cost matter more than the last percentage point. A one-sentence takeaway: Max when correctness is everything, High for everyday work.

The third is multi-token prediction, which lets the model predict several tokens in one forward pass instead of one at a time — faster inference, and better long-range coherence as a side effect.

Put together, the practical headline is the context window: up to 1,000,000 input tokens (via the glm-5.2[1m] identifier), with output up to 131,072 tokens. That's roughly five times GLM 5.1's ~200,000-token limit. A million tokens is enough to hold a mid-sized codebase in context at once — which is exactly the use case the whole design is pointed at.

How good is it, really

Here's where confidence layering matters, so I'll be explicit about what's a fact and what's a reported figure.

The fact: Z.ai shipped GLM 5.2 with no official benchmark suite. Every number you've seen circulating is either vendor-reported after the fact or from early independent evaluations, none of it broadly reproduced yet. Treat the specific decimals as directional, not gospel.

With that caveat, the reported figures are consistent across sources and point the same direction. On Terminal-Bench 2.1 (autonomous terminal-based coding), GLM 5.2 reportedly scores 81.0 — a large jump over GLM 5.1's 62.0, and within about four points of Claude Opus 4.8's 85.0. On SWE-bench Pro (resolving real software-engineering issues), it reportedly scores 62.1, ahead of GPT-5.5 at 58.6 and its own predecessor at 58.4, but behind Claude Opus 4.8 at 69.2. On Artificial Analysis's Intelligence Index it reportedly scored 51 — the highest of any open-weight model.

What gives those numbers more weight than the usual vendor table is independent confirmation that's harder to game. On Arena.ai's Code Arena — an Elo leaderboard built on blind, pairwise human votes — GLM 5.2 reportedly landed second overall. And on the crowdsourced Design Arena it reportedly took first place with an Elo of 1360, ahead of even Claude Fable 5. Blind human preference votes are much harder to manipulate than a self-reported pass rate, so those two results are the ones I'd trust most.

My read, stated as a judgment rather than a fact: GLM 5.2 is the strongest open-weight coding model available right now, it beats GPT-5.5 on several coding tasks, and it trails Claude Opus 4.8 on the hardest long-horizon work by somewhere between one and roughly thirteen points depending on the task. Close, not ahead — at a fraction of the price.

GLM 5.2 vs Claude Opus 4.8 vs GPT-5.5

For anyone choosing between the three, the trade-offs sort cleanly. The table is reported coding-benchmark scores plus the facts that don't move (pricing, context, licensing):

The honest summary: Claude Opus 4.8 is still the most capable of the three on the hardest agentic coding, and it's the safe default when correctness on long, autonomous runs is what you're paying for. GPT-5.5 sits in between on these particular coding benchmarks. GLM 5.2's case is not "it's the best" — it's "it's within a few points of the best, it's open, and it costs a fraction as much." If you're cost-sensitive, want to self-host, or want to fine-tune, that case is strong. If you're running mission-critical long-horizon agents where a few points of reliability pay for themselves, Claude Opus 4.8 is the more conservative pick. Pricing for the Claude side is published by Anthropic; the GLM figures are Z.ai's reported rates.

If you want to A/B the two closed rivals against your own prompts, both are callable through one API on GPT Proto — Claude Opus 4.8 (thinking) and GPT-5.5 — at a flat $4 per million tokens each. (That flat rate is GPT Proto's; the $5.00 / $25.00 input-then-output split in the table above is Anthropic's own list price for Opus 4.8 — same model, two different price structures.) Putting all three families behind a single key is the cheapest way to run the comparison yourself.

GLM 5.2 vs the GLM models you can use today

GLM 5.2 itself ships as open weights you download and host — Z.ai's hosted API is the only first-party way to call it, and as covered above that comes with a data-jurisdiction question. But the GLM line didn't start at 5.2, and the jump from the previous versions is the clearest way to see what actually changed.

The most useful comparison is against GLM 5.1, the immediate predecessor. Two differences stand out. The context window went from roughly 200,000 tokens to a full 1,000,000 — a five-fold jump that's the headline upgrade. And on coding, the reported gains are large: Terminal-Bench 2.1 climbed from 62.0 to 81.0, and SWE-bench Pro from 58.4 to 62.1. In other words, most of GLM 5.2's leaderboard standing is improvement over its own last release, not a small tweak.

If you'd rather call a hosted GLM through a single OpenAI-compatible API today rather than stand up the open weights, the GLM models GPT Proto currently carries are the ones just behind 5.2 in the lineage:

GLM-5.1 is the closest thing to 5.2 you can call here — same family, one generation back, with the ~200K context rather than 1M. For a lot of coding work that's a difference you won't notice; for repository-scale tasks that need the whole codebase in context at once, it's the gap that 5.2 closes. Full per-token rates for every model are on the model page.

Using GLM 5.2 in Claude Code, and a runnable example

One detail makes the GLM line unusually easy to drop into existing workflows: GLM 5.2 exposes an Anthropic-compatible endpoint. Tools built to talk to Claude — Claude Code, Cline, OpenCode — can point at it directly, swapping the model behind a coding agent without rewriting the integration. This is why "GLM 5.2 in Claude coding" is a real pattern and not just a search phrase: the agent harness stays the same, only the model underneath changes. (For 5.2 specifically that means Z.ai's own endpoint or a self-hosted deployment, since the open weights are the first-party route.)

If you'd rather not manage a deployment, the practical move today is to call a hosted GLM through GPT Proto's OpenAI-compatible API. Here it is against GLM 5.1 — the closest available sibling — which makes a good baseline before you decide whether 5.2's extra context is worth self-hosting:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_GPTPROTO_API_KEY",
    base_url="https://api.gptproto.com/v1",
)

resp = client.chat.completions.create(
    model="glm-5.1",
    messages=[
        {
            "role": "user",
            "content": (
                "Refactor this function for readability and explain the change:\n\n"
                "def f(x):\n"
                "    return [i for i in x if i % 2 == 0]"
            ),
        }
    ],
)

print(resp.choices[0].message.content)

The same request with cURL:

curl https://api.gptproto.com/v1/chat/completions \
  -H "Authorization: Bearer $GPTPROTO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.1",
    "messages": [
      {"role": "user", "content": "Write a Python function that returns the nth Fibonacci number, iteratively."}
    ]
  }'

Swap glm-5.1 for glm-5 or glm-5-turbo to trade quality for cost, or for claude-opus-4-8-thinking / gpt-5.5 to run the exact comparison from the table above — all through the same key.

You'll need that key first: create one from the GPT Proto dashboard, drop it into YOUR_GPTPROTO_API_KEY, and the call above runs as-is. Per-token rates for every model sit on the model page if you want to cost it out before committing.

Where it's strong, where it isn't

The strengths are concrete: it's the top open-weight coding model on the leaderboards that exist, it ships under a genuinely permissive MIT license, the million-token context is real and affordable to run thanks to IndexShare, and the cost-to-performance ratio is the best in its class.

The weaknesses are equally concrete, and worth stating plainly rather than burying. It trails Claude Opus 4.8 on the hardest long-horizon coding — the gap is small but consistent. Z.ai published no official benchmarks, so the numbers carry an asterisk until more independent labs reproduce them. And the cloud-API data-jurisdiction question is genuine: if your data can't legally or contractually leave a particular boundary, the hosted Z.ai API is the wrong door — self-host the open weights instead, which is the entire point of them being open.

Who should use it, and who shouldn't

Use GLM 5.2 if you're a developer who wants frontier-adjacent coding ability without frontier pricing, if you need to self-host or fine-tune, or if you're building a cost-sensitive agentic product where token spend dominates. It's an unusually good fit for anyone who already has a Claude-compatible agent harness and wants a cheaper engine behind it.

Reach for Claude Opus 4.8 instead if you're running mission-critical, long-horizon autonomous agents where the last few points of reliability are worth the premium, or if your work is bound by data-residency rules that the hosted GLM API can't satisfy and you can't self-host.