How to Use GLM 5.3: API Setup, Thinking Parameters & Coding Plan

Quick answer: GLM 5.3 is available three ways — the GLM Coding Plan (already rolled out to all subscribers), ZCode (Z.AI's coding agent), and the Z.AI API with model id glm-5.3. The one thing you must change: thinking.type can no longer be "disabled" — set it to enabled and pick reasoning_effort: low, high, or max.

TL;DR

Step 1: Choose How You Want to Use GLM 5.3

GLM 5.3 shipped on August 14, 2026 across three surfaces. Pick the one that matches your workflow:

A. GLM Coding Plan (easiest — already upgraded) All GLM Coding Plan subscribers were moved to GLM 5.3 on launch day. Nothing to install — your existing setup now runs 5.3. The plan moved to a points-based quota: input, cached input, and output tokens are metered separately. Key cost lever: calls outside peak hours cost 50% of standard points. Peak hours are 14:00–18:00 (UTC+8), Monday through Friday — everything else, including all weekend, is half price.

B. ZCode (best for agentic workflows) ZCode is Z.AI's coding agent, and it's where GLM 5.3's long-horizon strength shows best:

• 98%+ cache hit rate — repeated context bills at the lower cached rate (~30% more effective tokens).
• 1.5× limited-time quota boost through August 31 — stackable with cache savings for up to 180% of standard quota.
• Goal mode — plans, codes, tests, and verifies until the target is met (this is the long-horizon RL payoff).
• Remote Control — monitor and steer long-running tasks from your phone via WeChat or Feishu.

C. Z.AI API (for your own apps) Model id glm-5.3, same endpoint family as 5.2. One required change (next step).

Step 2: Fix Your Thinking Parameters (Required)

GLM 5.3 no longer supports disabling thinking. If your application currently sends thinking.type: "disabled", the request will fail once you switch to glm-5.3.

The new minimum request:

{
  "model": "glm-5.3",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "max"
}

Thinking parameters:

Migration order (to avoid failed calls):

1. Keep your current model ID.
2. Change thinking.type to "enabled".
3. Set reasoning_effort: "low" as a safe starting point.
4. Update the model ID to glm-5.3.
5. For coding workloads, raise reasoning_effort to max — Z.AI's recommendation.

Step 3: Optimize Cost

GLM 5.3 is more token-efficient than 5.2 (better results at fewer output tokens on Z.AI's code bench), but you can push that further:

• Schedule heavy jobs off-peak. 50% points outside 14:00–18:00 UTC+8 weekdays. Batch big runs for evenings/weekends.
• Exploit caching. Long, stable system prompts and context prefixes hit the cached-input rate. In ZCode, cache hits give ~30% more effective tokens.
• Stack the August boost. ZCode's 1.5× quota boost runs through August 31 and stacks with cache savings.
• Right-size effort. Don't run max on trivial requests. Use low for classification/extraction, high for ordinary coding, max for complex multi-file work.

Step 4: Run It in Your Coding Agent

GLM 5.3 works with ZCode, Claude Code, OpenCode, and more via Z.AI's devpack (docs.z.ai/devpack/overview). The setup pattern:

1. Install the devpack/configure the provider for your agent.
2. Point the model at glm-5.3 (or the Coding Plan backend if subscribed).
3. Set reasoning_effort: "max" in the agent config for repo-scale tasks.
4. For multi-hour jobs, use ZCode's Goal mode so the agent verifies completion rather than stopping at first draft.

Step 5: Self-Host (Coming in Two Weeks)

GLM 5.3 weights will be public on HuggingFace two weeks after launch — Z.AI is holding them for safety evaluation and hardening, especially because of the model's emergent cybersecurity capability. If you plan to self-host, that's the timeline to watch.

What Changed From GLM 5.2 — Quick Reference

FAQ

Is GLM 5.3 available in the API? Yes — model id glm-5.3, per-token billing. Update your thinking parameters first.

Do I need to change my code from GLM 5.2? Only if you used thinking.type: "disabled" — set it to enabled with reasoning_effort: "low" before switching model IDs.

Is GLM 5.3 free? GLM Coding Plan subscribers get it immediately; off-peak usage is 50% of standard points. No separate free tier announced.

Which reasoning effort should I use for coding? max — Z.AI's explicit recommendation for coding tasks. Use low for lightweight extraction/classification.

Can I run GLM 5.3 locally? Soon — weights release on HuggingFace two weeks after the August 14 launch.

Does GLM 5.3 work with Claude Code / OpenCode? Yes, via Z.AI's devpack, along with ZCode.

Sources

• Z.AI: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities (August 14, 2026)
• Z.AI DevPack Overview

Last updated: August 14, 2026