Migrate GLM-5.3 thinking parameters

Core content summary

The official Z.ai documentation for the GLM-5.3 text model covers its capabilities, feature changes, parameter descriptions, and integration details—the first-hand authoritative reference for prompt design and API calls.

Overview

• GLM-5.3 is Z.ai's latest flagship model, with major advances across complex software engineering and Agent tasks. It uses the same base model as GLM-5.2; all improvements come from post-training.

• Stronger coding: Z.ai's internal Z.ai Code Bench score is up 50% versus GLM-5.2; Terminal Bench 3.0 and Agents' Last Exam (CLI) are open-source SOTA.

• Emergent cybersecurity capabilities: CyberGym is currently best in vulnerability discovery; the deeper the exploitation chain, the more significant the gains (the vulnerability-exploitation benchmark score is more than twice that of 5.2).

Key feature changes

• Text modality only; a 1M-token context window and a maximum output of 128K tokens.

• Thinking is always enabled, with three reasoning-intensity levels: low, high, and max; disabling thinking is no longer supported.

• Migration note: If an application currently uses thinking.type: "disabled", change it to enabled and set reasoning_effort to low before updating the model ID to glm-5.3. Otherwise, the request will fail.

• For complex tasks such as coding, max is recommended. Example:

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

How to use

• GLM Coding Plan: Fully available, with GLM-5.3 supported in commonly used coding agents. The new version uses a credit-based quota system; calls during off-peak periods (including all day on weekends) consume only 50% of the standard credits.

• Model API: Coming soon; supported protocols and integration endpoints:

  • OpenAI Chat Completion protocol: https://open.bigmodel.cn/api/paas/v4

  • OpenAI Response protocol: https://open.bigmodel.cn/api/v1

  • Anthropic Message protocol: https://open.bigmodel.cn/api/anthropic

• Supported capabilities: thinking mode, streaming output, Function Calling (tool calls), context caching, and structured output (JSON and more).

Official capability details (for prompt design reference)

• The training environment shifted from traditional programming problems to "complete professional work units": some tasks take a senior engineer several days to complete (for example, an ML infrastructure task: identify a bottleneck in the training stack → implement an optimization → run experiments → deliver measurable end-to-end acceleration).

• Validators are generated without reference solutions and are used for training only after checks with an oracle, no-op, or unresolved state; the RL strategy of SAO with compaction continues.

• Z.ai Code Bench: 34.5% at the Max tier with approximately 75,000 tokens (5.2 was 23.4% at approximately 96,000); 31.4% at the High tier with approximately 50,000 tokens, exceeding Opus 4.8 (29.5% at approximately 120,000); Fable 5 still leads at 39.5% at the Max tier.

• Cybersecurity: CyberGym 84.5% (best, ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%); ExploitBench 54.4% (5.2 was 24.4%, more than doubled); ExploitGym 2h/6h = 105/130; 2,436 vulnerabilities across 269 real-world codebases (1,097 medium- or high-severity), with the oldest vulnerability remaining latent for approximately 40 years.

Prompt design takeaways (based on official guidance)

1. Thinking cannot be disabled: There is no need to tell the prompt to "stop thinking"; use reasoning_effort to select a tier instead—low for simple tasks to save tokens, and max for complex coding or Agent tasks.

2. Treat the model as a "structured engineering model," not a conversational chatbot: Describe tasks, data, and constraints separately.

3. Make good use of tool calls and structured output: The official service supports Function Calling, JSON structured output, and context caching; prompts can declare schemas and tools.

4. When migrating an existing application, change thinking.type and reasoning_effort first, then switch the model ID.