GLM-5.2 Official Documentation: Overview and API Quick Start (docs.z.ai)
The official standard integration configuration for GLM-5.2 is: model name `glm-5.2`, a 1M context window / 128K maximum output, `thinking.type: enabled` + `reasoning_effort: max`, and `temperature: 1.0`. You can copy the curl / Python examples directly to make your first call and review the typical use cases identified by the official documentation..
Official Configuration Guide for Migrating from GLM-5.1 / GLM-5 / GLM-4.x to GLM-5.2
The official GLM-5.2 migration checklist and parameter configuration: change the model ID to `glm-5.2`; use the default `temperature` of 1.0 or default `top_p` of 0.95 (tune only one of the two); enable thinking by default; use `high` or `max` for `reasoning_effort`; configure streaming and streaming tool calls (`stream=true` + `tool_stream=true`) as specified by the official guidance; and use the included Python migration example directly..
Prepare
old request, new model ID, streaming flags, regression task
The official documentation states that thinking is enabled by default for GLM-5.2 (as with GLM-5.1/5/4.7), and provides four thinking modes: default thinking, interleaved thinking (thinking between tool calls), preserved thinking (retaining reasoning content across turns with `clear_thinking: false`), and turn-level thinking (an independent switch for each turn). It also highlights a key constraint for Agent integrations: historical `reasoning_content` must be returned unchanged..
Prepare
reasoning_content history, clear_thinking, turn policy, tool result
Using GLM-5.2 (zai-glm-5-2) Through Mistral: Third-Party Hosting Configuration and Pricing
Mistral now hosts GLM-5.2 as a third-party open model (Public Preview, model ID `zai-glm-5-2`, 1M context / 128k output, with no modifications), so it can be accessed directly across the Mistral ecosystem (including Vibe CLI) using that ID, at $1.4 / $0.14 (cached input) / $4.4 (output) per million tokens..
Prepare
Mistral model ID, billing tier, context workload, smoke call
GLM-5.2 Official Release Notes and Complete Benchmark Table (Z.ai Blog)
Z.ai’s 2026-06-16 release positions GLM-5.2 as a 1M-context long-horizon flagship and reports 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-Bench Pro; it also discloses training-stage reward-hacking risk.
Evidence
Vendor report
Boundary
Does not make official scores stable across providers or a safety guarantee.
NIST (National Institute of Standards and Technology) official news siteEditorial analysis
NIST CAISI's Independent Capability Assessment of Z.ai GLM-5.2
NIST CAISI published its assessment on 2026-07-17 after completing it on 2026-07-08: GLM-5.2 was similar to GPT-5.2 overall and Opus 4.6 on cyber capability, while safeguards were mixed for agentic exploits and biological questions.
Evidence
Editorial analysis
Boundary
Does not treat self-hosted open-weight safety as the same evaluated condition.
Semgrep IDOR Benchmark: GLM-5.2 Results with a Prompt-Only Setup in Security Code Auditing
Semgrep’s 2026-06-22 IDOR benchmark held dataset, evaluation, and prompt constant: GLM-5.2 reached 39% F1 in a Pydantic AI prompt-only harness at about $0.17 per vulnerability; this is not a general cyber score.
Evidence
Editorial analysis
Boundary
Does not generalize one IDOR result to SSRF, production audits, or all repositories.
