2026-06-18站内改写5 min readUpdated: 2026-06-18

An open-source AI just beat OpenAI's GPT-5.5 at coding (1/6th the price)

GLM-5.2, an open-source model by Z.AI, achieves top coding performance with 1M context, outperforming GPT-5.5 on several benchmarks at a fraction of the cost.

SourceHacker News AIAuthor: Raj_Sidwadkar

Overview

GLM-5.2 is a flagship model built for the era of long-horizon tasks. With truly usable 1M-token context, it has been tested to handle project-scale engineering context, delivering more stable long-task execution, more reliable adherence to engineering standards, and higher success rates in development scenarios. A single task can complete the full development workflow—from requirements to deployable products across multiple platforms.

Positioning

Flagship Foundation Model

Input Modalities

Text

Output Modalitie

Text

Context Length

Maximum Output Tokens

128K

Capability

Thinking Mode

Offering multiple thinking modes for different scenarios

Streaming Output

Support real-time streaming responses to enhance user interaction experience

Function Call

Powerful tool invocation capabilities, enabling integration with various external toolsets

Context Caching

Intelligent caching mechanism to optimize performance in long conversations

Structured Output

Support for structured output formats like JSON, facilitating system integration

MCP

Flexibly integrate external MCP tools and data sources to expand application scenarios

Usage

Project-Level Codebase Takeover: Let the Model Understand an Entire Project in One Go

This is the best starting point to experience the generational leap of GLM-5.2. It can continuously retain module boundaries, architectural constraints, API contracts, directory structures, and historical decisions, significantly reducing the sense of context fragmentation in the later stages of long-running tasks. For complex projects, the key experience is that the model does not merely read more context—it can carry forward the engineering judgments formed earlier into subsequent execution.Recommended way to try it: Choose a real business codebase, preferably one that includes backend, frontend or client-side code, configuration files, tests, documentation, and engineering conventions. First, ask the model to perform a technical audit:

Please read the current project and output a system architecture map, core module responsibilities, key API contracts, major data flows, core call chains, potential technical debt, and the engineering constraints that must be followed in future refactoring.

Long-Horizon Refactoring: Let It Run a Real Engineering Task End to End

GLM-5.2 is more stable in cross-file, multi-step, long-chain tasks. It first breaks down the goal, identifies dependencies and risks, then implements, verifies, and closes the task in stages. This makes it suitable for tasks that require continuous progress, such as module decoupling, API migration, directory restructuring, SDK adaptation, and cross-language refactoring.Recommended way to try it: Choose a medium-sized refactoring task, define clear boundaries, and enable /goal mode:

Please complete the decoupling and refactoring of the current module without changing the business logic, API signatures, or runtime behavior. First provide the execution plan, impact scope, risk boundaries, and verification method. After completion, run the necessary tests and output the verification results.

Production-Grade Standards Stress Test: See Whether It Can Hold the Line on Hard Engineering Constraints

GLM-5.2 shows stronger consistency in following engineering standards, especially in long-context and multi-round execution. It is better at adhering to code style, architectural boundaries, dependency constraints, build processes, testing requirements, and commit boundaries, reducing risks such as out-of-scope changes, invalid dependencies, skipped verification, or unauthorized commits.Recommended way to try it: Provide the model with your team’s real engineering standards, such as lint rules, build commands, testing requirements, commit conventions, and prohibited actions in CLAUDE.md or Agent.md. Then give it a real modification task:

Please strictly follow the engineering standards of the current repository. Do not introduce new dependencies, do not modify API contracts, and do not commit changes proactively. After completing the modification, run the build, lint, and tests, then report the verification results and any uncovered risks.

Mobile On-Device Debugging Loop: From Code Implementation to Device Validation

In mobile development scenarios, GLM-5.2 can cover client-side architecture, streaming messages, long-connection states, local state management, keyboard behavior, scrolling logic, system notifications, permission mechanisms, and background recovery. More importantly, it can use ADB, logcat, screenshots, and runtime logs to locate real-device issues, making it closely aligned with practical mobile engineering workflows.Recommended way to try it: Choose a real Android or Mini Program task and let the model go from implementation to validation:

Please implement a native Android client in Kotlin that connects to the existing server-side API and supports multi-session conversations, streaming messages, voice input, notifications, and reconnection after disconnection. After completion, install it on a real device using ADB, and debug it with logcat and screenshots.

WeChat Mini Program Development: Migrating from a Web App to a WeChat Mini Program

Mini Game Development: From Gameplay Rules to a Playable Loop

GLM-5.2 is well suited for testing rule understanding, state machine design, level structure, scoring logic, resource loading, interaction feedback, and settlement flows in mini game development. Compared with static pages, this type of task better demonstrates the model’s understanding of complex states, user paths, and product completeness.Recommended way to try it: Provide a complete but not overly detailed gameplay goal, and let the model first design the rules, then implement a runnable version:

Please develop a lightweight level-based mini game. First design the core gameplay loop, state machine, level structure, scoring rules, failure and settlement logic, then implement basic features including start, pause, resume, settlement, restart, and local save. After completion, explain the project structure, verified features, and possible next-step extensions.

Research Reproduction: From Paper and Data to a Runnable Engineering Project

GLM-5.2 can turn the model architecture, loss functions, data pipelines, and training/inference scripts described in a paper into runnable code that aligns with the paper. It can correctly set up the model structure in one pass, maintain consistency across multiple files, and autonomously run, debug, and fix code and environment issues. What it delivers is not just code snippets, but an engineering project that can truly reproduce the paper’s reported results.Recommended way to try it: Pick a paper with a model and experiments, preferably one with open-source code or public metrics, and provide the paper and data to the model. See whether it can implement the model, run it successfully, and align the results with the paper:

Please reproduce the experiments based on this paper and dataset. Fill in implementation details not explicitly described in the paper. Use PyTorch to build the model architecture and loss functions, construct the data pipeline and training/inference scripts, and ensure the project runs successfully with consistency across multiple files. Autonomously identify and fix runtime issues, verify the paper’s metrics item by item until they are aligned, and explain the reproduction path, key changes, and any remaining gaps.

Code-to-Video Loop: From Natural-Language Ideas to a Demo-Ready Video

In Code-to-Video scenarios, GLM-5.2 can use the Remotion framework to create videos programmatically with React code, including components, parameters, and animation logic, and then render them into MP4. In simple terms, it treats video creation as writing code. It covers the full workflow from translating natural-language ideas into Remotion React code to rendering video output, enabling code-driven generation of a runnable, demo-ready video.Recommended way to try it: Choose a real video creation task and let the model start from a single natural-language idea, then gradually produce a renderable, playable, and iterable video:

Please create a new composition in Remotion and add a map. Start from Los Angeles, zoom the camera out while keeping LA in focus. Then draw an animated route from Los Angeles to New York and have the camera follow the route. Add one more stop to the journey — this time, we are going to Paris.

Introducing GLM-5.2

1M Context: Making Long-Horizon Tasks Stable and Practical

The foundation of long-horizon tasks is not having a 1M context, but making 1M context truly usable. GLM-5.2 delivers a Solid 1M lossless context and has undergone months of specialized training for long-horizon Coding Agent scenarios, covering high-value tasks such as large-scale implementation, automated research, and performance optimization.Compared to solutions that merely extend context length, GLM-5.2 maintains more stable performance at ultra-long context, even surpassing Opus in select real-world benchmarks.GLM-5.2 delivers state-of-the-art long-horizon coding performance among open-source models. Across FrontierSWE, PostTrainBench, and SWE-Marathon, it consistently ranks among the top models overall—trailing Opus 4.8 by just 1% on FrontierSWE, outperforming GPT-5.5 and Opus 4.7 on multiple benchmarks, and remaining the highest-ranked open-source model across all three. These results demonstrate that GLM-5.2’s 1M context window translates into practical long-horizon engineering capability.

Coding Capabilities Validated by Both Benchmarks and Developers

On standard coding benchmarks, GLM-5.2 is the strongest open-source model, improving on GLM-5.1 by a wide margin: 81.0 vs. 62.0 on Terminal-Bench 2.1 and 62.1 vs. 58.4 on SWE-bench Pro. It also closes much of the gap to the closed-source frontier — on Terminal-Bench 2.1 (81.0) it lands within a few points of Claude Opus 4.8 (85.0) — while staying ahead of Gemini 3.1 Pro. Before its official release, GLM-5.2 was made available in advance to GLM Coding Plan users. Developers reported improvements mainly in the following areas:

Stronger project-level context capacity, enabling an entire codebase to be placed within a single reasoning workflow;

More stable long-horizon task execution, allowing complex tasks to progress continuously without easily going off track;

More reliable adherence to production-grade engineering standards, helping enforce hard constraints in team development workflows;

Stronger client-side and mobile engineering capabilities, going beyond app generation to support a complete on-device debugging loop.

↗ Blog

Resources

API Documentation: Learn how to call the API.

Quick Start

The following is a full sample code to help you onboard GLM-5.2 with ease.

cURL

Official Python SDK

Official Java SDK

OpenAI Python SDK

Basic Call

curl -X POST "https://api.z.ai/api/paas/v4/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer your-api-key" \ -d '{ "model": "glm-5.2", "messages": [ { "role": "system", "content": "You are a senior full-stack software engineer, proficient in frontend development, backend architecture design, and modern web technology stacks." }, { "role": "user", "content": "Design and build a personal blog website for me, including a homepage, article list page, and article detail page, using React + Node.js technology stack." } ], "thinking": { "type": "enabled" }, "reasoning_effort": "max", "max_tokens": 4096, "temperature": 1.0 }'

Streaming Call

[truncated for AI cost control]