GLM-5 Review: GLM-5.2, 5.1 & 5 — Benchmarks, Pricing & Access

GLM-5 is the model that put Zhipu AI (Z.ai) on the frontier — and its newest version, GLM-5.2, is the top-ranked open-weights model on the independent Artificial Analysis Intelligence Index. It's MIT-licensed, has a 1M-token context window, and costs a fraction of the Western frontier. This review maps the fast-moving GLM-5 line (GLM-5 → 5.1 → 5.2), separates the independently-verified benchmarks from the vendor-reported ones, gives the real Z.ai pricing, models what it costs to run, and shows how to call it.

What GLM-5 is

GLM-5 is Zhipu AI / Z.ai's frontier-class model family, built around the thesis of its technical paper — "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763). All three current versions are open-weights, MIT-licensed, text-only Mixture-of-Experts LLMs (~744-753B total parameters, ~40B active per token), available on Hugging Face under zai-org and via the Z.ai API. The newest, GLM-5.2, is the top-ranked open-weights model on the independent Artificial Analysis Intelligence Index — a genuine frontier-adjacent model at open-weights prices.

The version map (5 → 5.1 → 5.2)

The cadence is fast — three releases in four months — and all stay open-weights (MIT) with FP8 variants. GLM-5.1 shipped proprietary-first, then open-sourced around April 8; GLM-5.2's subscriber rollout began ~June 13 before the June 16 standalone API.

Benchmarks (verified vs vendor)

It's worth separating what's independently verified from what the vendor reports:

• Independent (Artificial Analysis): GLM-5.2 scored 51 on the Intelligence Index (v4.1) — #1 among open-weights models at release, and #4 overall (behind Claude Fable 5, Opus 4.8, and GPT-5.5). This is the strongest verified signal.
• Vendor-reported (Hugging Face model card): SWE-bench Pro 62.1, AIME 2026 99.2, GPQA-Diamond 91.2, Terminal-Bench 2.1 ~81-82.7. Strong numbers, but not independently reproduced.
• GLM-5.1: vendor-reported SWE-bench Pro 58.4 and an up-to-8-hour autonomous single-task claim.

The honest read: GLM-5.2 is a legitimately frontier-adjacent model — independently the best open-weights option — with vendor benchmarks that look frontier-competitive but should be treated as vendor framing until reproduced.

How GLM-5.2 did in our test

Beyond the leaderboards, we ran GLM-5.2 ourselves through our executed-code coding benchmark — nine tasks scored by running the generated code against hidden tests (see our methodology). It scored 9/9 — a clean sweep — competitive with the frontier on correctness. The cost story is more nuanced than the sticker price: GLM-5.2 reasons moderately, emitting ~560 reasoning tokens per task and averaging ~12 seconds per task. That put its real billed cost at about $1.99 per 1,000 tasks — below Claude Opus 4.8's $4.05, but well above lean coders like DeepSeek V4-Flash ($0.13) and Qwen3 Coder Next ($0.10).

Pricing

GLM's prices are a fraction of the Western frontier — GLM-5.2's combined per-token cost (~$5.80/1M) is roughly one-sixth of GPT-5.5's:

Z.ai API rates (per 1M tokens): GLM-5.2 $1.40 in / $4.40 out ($0.26 cached); GLM-5.1 $1.40 / $4.40; GLM-5 $1.00 / $3.20 ($0.20 cached). And because the weights are MIT-licensed, self-hosting is an option if you have the GPUs.

Note: those are Z.ai's own API rates. Through OpenRouter and gateways, GLM-5.2 is often cheaper — around $0.95 in / $3.00 out per 1M (GLM-5 ~$0.60 / $1.92) — so price it on the endpoint you'll actually call.

What GLM-5.2 costs to run

Modeled monthly cost across five workloads — GLM-5.2 sits well below the Western flagship, with GLM-5 cheaper and DeepSeek cheaper still:

How to access it

Three ways: the Z.ai API, a gateway, or self-host. The API is OpenAI-SDK-compatible:

from openai import OpenAI
client = OpenAI(base_url="https://api.z.ai/api/paas/v4", api_key="$ZAI_API_KEY")
resp = client.chat.completions.create(
    model="glm-5.2",                      # or glm-5.1, glm-5
    messages=[{"role": "user", "content": "Plan and implement this feature..."}],
)
print(resp.choices[0].message.content)

Or reach GLM-5.2 through a gateway with one key alongside 300+ other models; or download the MIT weights from Hugging Face (zai-org/GLM-5.2) to self-host.

Who should use GLM-5.2

FAQ