Notice: GPT Image 2.6 Flare and Sunburst are newly launched with a limited-time 20% off, try it now
HiAPI
glm-5.3 详情页 Hero 图:左侧奶油色柔光留白,右侧悬浮半透明玻璃文档层叠与光带轨道,冰蓝、腮红粉、薰衣草蓝渐变,发光立体字 glm-5.3 嵌于玻璃面板。

Zhipu AI

GLM 5.3

glm-5.3

Zhipu's flagship text reasoning model, focused on complex software engineering and long-horizon Agent tasks, with roughly 1M context, always-on reasoning and support for multiple reasoning-effort levels, delivering more token-efficient output at the same score

glm-5.3
Input
Max output
131,072 tokens
Release date
2026-08-14
Streaming Structured JSON

Pricing

glm-5.3 /1M tokens

Specifications Unit Price
Input $0.85479
Output $2.99178

Overview

Positioning Overview

glm-5.3 is a text reasoning model released by Zhipu, positioned for complex software engineering and long-horizon Agent tasks. It builds on the previous-generation base model, with all improvements coming from post-training. Reasoning is always enabled and supports three intensity levels: low, high, and max, with a context window of approximately 1 million tokens. It is suitable for developers who need the model to autonomously complete multi-step tool calls, cross-file code understanding, and terminal operations, as well as for teams focused on reasoning token efficiency who want to reduce output costs while achieving the same task scores.

Core Highlights

  • Significantly improved coding and Agent capabilities over the previous generation — officially reported to improve by about 50% compared to GLM-5.2 on the proprietary Z.ai Code Bench, and achieves open-source SOTA on public benchmarks such as Terminal Bench 3.0.
  • Fewer output tokens at higher scores — the Max level uses about 75K tokens per task compared to 96K for the previous generation, with simultaneous improvements in performance and token efficiency, making long-task costs more controllable.
  • Three adjustable reasoning intensity levels per task — low, high, and max correspond to different reasoning investments, with max as the default, making it easy to trade off between latency and quality.
  • Approximately 1M token context window — can incorporate large codebases and long conversations at once, supporting holistic analysis across files and modules.

Capability Description

Ultra-long context understanding: The context window is approximately 1,048,576 tokens, allowing large codebases or multi-turn long conversations to be placed into a single request, reducing the engineering burden of chunking and concatenation.

Adjustable reasoning intensity: Reasoning is always enabled and supports three intensity levels: low, high, and max, with max as the default; it can be selected based on task complexity to balance response speed and reasoning depth.

Software engineering and Agentic Coding: Designed for multi-step tool calls and code generation/refactoring, officially reported to improve by about 50% over the previous generation on Z.ai Code Bench, suitable for letting the model autonomously drive complex engineering tasks.

Terminal and long-horizon task execution: Achieves open-source SOTA on benchmarks such as Terminal Bench 3.0 and Agents' Last Exam, suitable for command-line operation automation and Agent workflows requiring continuous multi-turn decision-making.

Text reasoning and output: Uses text as the input and output modality, billed by token, and can be directly integrated into conversational, analytical, and generative business pipelines.

Applicable Scenarios

  • Software teams analyzing large codebases — input cross-file source code and requirement descriptions, produce refactoring plans or change suggestions, for use in daily development and code review processes.
  • Agent developers building multi-step tool-call workflows — input task goals and tool definitions, produce continuous multi-turn call decisions, for use in automated workflows and background task orchestration.
  • Operations and toolchain teams automating terminal tasks — input command environments and task instructions, produce executable command-line operation sequences, for use in script generation and terminal operation automation.
  • Security teams evaluating vulnerability discovery — input target code and evaluation conditions, produce vulnerability leads and analysis conclusions, for use in internal security assessments; dual-use risks must be evaluated first.

Limitations and Selection Advice

The reasoning process cannot be disabled, and even low-complexity requests incur additional reasoning latency and cost; lighter models can be used for scenarios such as simple classification and short-text rewriting. Capability improvements come from post-training rather than architectural iteration, and the base model is the same as the previous generation. Capabilities related to vulnerability discovery and exploit chains have grown significantly, and dual-use risks should be independently assessed before use in security scenarios. Model weights are released later, with API calls being the primary method initially. Billing is measured by token, and long contexts and max-level reasoning will significantly amplify consumption.

  • Price update glm-5.3 price adjustment

Related Models

View all models →

Same Type