Zhipu AI

GLM-5.2

Provider

Zhipu AI

Status

available

Context

1,000,000 tok

SWE-bench

62.1%

Price

$1.2 / $4.1 /MTok

GLM-5.2 is Zhipu AI (Z.ai)‘s open-weight flagship and the current release of the GLM-5 series, shipped on 13 June 2026 under the permissive MIT licence. It is a large mixture-of-experts model (~753B total parameters) with a 1-million-token context window, built specifically for long-horizon coding agents — Zhipu says it beats GPT-5.5 on several multi-step coding benchmarks at roughly a sixth of the cost (VentureBeat).

It is the third release in a fast cadence — GLM-5 (February 2026), GLM-5.1 (April 2026) and now GLM-5.2 — each sharpening Zhipu’s “agentic engineering” focus. The line sits in the Chinese open-weight cluster with DeepSeek V4, Kimi K2.6 and MiniMax M3; Zhipu’s coding benchmarks rank GLM at or near the top of the open field, while independent indices place it competitively but below the open leaders on the hardest general tasks.

Quick specs

VIEW GLM-5.2 →

What GLM-5.2 is

GLM-5.2 is a mixture-of-experts model — about 753 billion total parameters with ~40 billion active per token — using sparse attention (Zhipu’s “IndexShare” reuses indexers across sparse-attention layers to cut per-token compute at long context) (Hugging Face). It supports a 1-million-token context and up to 128K output tokens, with configurable thinking effort (up to a max reasoning level), native tool calling, MCP integration and structured JSON output (Z.ai docs).

Its purpose is explicit: long-horizon coding agents. Zhipu describes months of specialised training for agentic software-engineering scenarios, and markets the series’ ability to sustain autonomous work across hundreds of iterations without the early plateau seen in prior models. It is text-in, text-out, and — like the rest of the GLM-5 line — released as MIT open weights, free to self-host. Notably, GLM-5 was reportedly trained on Chinese accelerators (Huawei Ascend and others) rather than Nvidia (Reuters via AOL).

The GLM-5 lineage

All three are open-weight and MIT-licensed; the jump to a 1M-token context is GLM-5.2’s headline change, alongside higher coding-benchmark scores.

Benchmark performance

Zhipu’s vendor figures are strong on coding and agentic tasks; independent indices are more conservative.

The pattern is a coding- and agent-focused open model that ranks at or near the top of the open field on Zhipu’s own benchmarks, while the independent Artificial Analysis index put the predecessor GLM-5.1 around the middle of the open pack — behind Kimi K2.6 and DeepSeek V4-Pro on its composite. One caveat worth stating plainly: the widely-cited “GLM ahead of DeepSeek, behind Kimi” SWE-bench Pro ranking comes from third-party trackers, not from NIST’s CAISI evaluation (which assessed only DeepSeek V4-Pro). See best AI for coding for cross-model standings.

Pricing and access

The open weights are free to download and self-host under MIT, from Hugging Face (with FP8 and GGUF quantisations), via vLLM, SGLang, Transformers, Ollama and LM Studio. Hosted, GLM-5.2 is about $1.20 input / $4.10 output per million tokens on OpenRouter; GLM-5.1 lists at $1.40 / $4.40 with cached input at $0.26 on Z.ai’s own API (Artificial Analysis). VentureBeat estimated GLM-5.2’s cost at roughly one-sixth of GPT-5.5 for comparable long-horizon coding. Zhipu also sells a flat-rate GLM Coding Plan that plugs into agentic IDE tools. The API model ID is glm-5.2, served from the Z.ai / BigModel platform.

One access caveat: hosted use routes through Chinese infrastructure, and Zhipu is on the US Entity List — relevant for data-residency, compliance and government procurement. Self-hosting the open weights avoids the hosted-API concerns.

How GLM-5.2 compares

• vs other open models — In the Chinese open-weight cluster with DeepSeek V4, Kimi K2.6 and Qwen, GLM’s draws are its long-horizon agentic-coding tuning, a 1M context (5.2) and a developer-friendly coding subscription. On the independent intelligence index it trails Kimi K2.6 and DeepSeek V4-Pro on the hardest general tasks.
• vs the proprietary frontier — Zhipu claims GLM-5.2 beats GPT-5.5 on several long-horizon coding benchmarks at a fraction of the price; the closed leaders (Claude Opus 4.8, GPT-5.5, Gemini 3.x) still lead on the hardest general reasoning and on independent composites.
• within the GLM-5 series — GLM-5.2 adds the 1M context and higher coding scores over GLM-5.1 (200K); GLM-5 was the first of the line.

Known limitations

Below the closed frontier on the hardest general-reasoning tasks, and behind the top open models on independent composites. Vendor-heavy benchmarks — the headline “beats GPT-5.5 / Opus” claims are Zhipu-reported and need independent corroboration; the cross-model SWE-bench Pro ranking comes from aggregators, not CAISI. China-hosted API and Entity List — hosted use carries data-residency, compliance and procurement considerations (self-hosting mitigates this). Text-only — no verified vision/image input despite a multimodal architecture class in the model metadata.

FAQ

What is GLM-5.2?

GLM-5.2 is Zhipu AI (Z.ai)‘s open-weight flagship, released 13 June 2026 — a ~753B-parameter mixture-of-experts model under the MIT licence, with a 1-million-token context window, built for long-horizon coding agents.

Is GLM-5.2 open source?

Yes — it is released as open weights under the MIT licence, free to download, self-host, fine-tune and use commercially, from Hugging Face. Zhipu monetises through its hosted API and the GLM Coding Plan rather than by withholding the weights.

How much does GLM-5.2 cost?

About $1.20 per million input tokens and $4.10 per million output on OpenRouter — roughly a sixth of GPT-5.5’s cost for comparable long-horizon coding, by VentureBeat’s estimate. The open weights are free to self-host.

How good is GLM-5.2 at coding?

Strong: Zhipu reports 62.1% on SWE-bench Pro and 81.0% on Terminal-Bench 2.1, claiming it beats GPT-5.5 on several long-horizon coding benchmarks. These are vendor figures; on independent composites the line rates competitively but below the top open models.

What is the difference between GLM-5, GLM-5.1 and GLM-5.2?

They are successive releases in the GLM-5 series: GLM-5 (Feb 2026) and GLM-5.1 (Apr 2026) had ~200K context, while GLM-5.2 (Jun 2026) extends the context to 1M tokens and raises the coding-benchmark scores. All are open-weight and MIT-licensed.

Is GLM-5.2 safe to use for sensitive data?

The open weights are safe to self-host (nothing leaves your infrastructure). The hosted Z.ai API routes data through Chinese infrastructure and the company is on the US Entity List, so for sensitive, regulated or government work prefer self-hosting or a Western-hosted provider such as Anthropic or Mistral.

Last verified 19 June 2026. The MIT licence, MoE architecture, 1M context and the GLM-5 → 5.1 → 5.2 lineage are confirmed via Z.ai’s docs, Hugging Face model cards, Artificial Analysis and OpenRouter. Benchmark figures are largely vendor-reported (Zhipu’s own testing) and run higher than independent indices; the cross-model SWE-bench Pro comparison is from third-party trackers, not NIST CAISI. Confirm against current leaderboards before relying on them.