Zhipu AI

GLM-5.2

Released Jun 13, 2026 · knowledge cutoff 2026-03

Status: Superseded
Location: China
Type: LLM
Modality: Text
Context window: 1M
Max output: 128K
Speed (max effort): 69 tok/s · 1.61s to first answer token
Price ($/MTok in / out): $1.40 / $4.40
Cost per task (max effort): $0.30
Open weights: Yes

Benchmarks8 of 10 reported

MMLU-Proretired

80.6%

GPQA Diamond

91.2%

SWE-bench Pro

62.1%

Terminal-Bench 2.1

81%

AIME (latest)retired

99.2%

Humanity's Last Exam

40.5%

LMArena Elo

1469

ARC-AGI-2

22.8%

SWE-bench Verified

—

GDPval-AA v2

—

Strengths & weaknesses

Strengths

Produces clean, production-ready code first time more often than its cost suggests
Reliable multi-turn tool calling — it doesn't lose track of what it already called
Unrestricted licence and mainstream inference-stack support make self-hosting straightforward

Weaknesses

Very verbose — roughly double its predecessor's output per task, which shows up on both the bill and the clock
Noticeably slower per task than closed flagships of the same period
Text-only, so no screenshots, diagrams or PDFs

SWE-bench Pro 62.1 reported; no SWE-bench Verified figure has been published. GPQA Diamond sources conflict (91.2 per the official Hugging Face model card vs ~80.3 in some third-party trackers) — official figure used. HLE 40.5% no tools / 54.7% with tools. Max output listed as 128K in most reporting; some sources cite up to 163,840 for extended reasoning. LMArena Elo 1469 is the "GLM-5.2 (max)" text-arena listing (arena.ai), the only GLM-5.2 variant on the board. Terminal-Bench 2.1 81.0 is Z.ai's own developer-docs figure (reproduced by DataCamp, Semgrep and BenchLM's TB 2.1 board); an open GitHub issue (zai-org/GLM-5#100) reports difficulty reproducing it, so treat it as self-reported.

For developers

API model strings

Not researched

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor