ALL INTELLIGENCE
智谱 AI 研究T1·Agents

Zhipu Releases GLM-5-Turbo: Optimized for OpenClaw Agent Tasks

By 智谱 AIOriginal title: GLM-5-Turbo:龙虾增强的基座模型
68AI Score

Zhipu AI released GLM-5-Turbo, a base model deeply optimized for the OpenClaw agent ecosystem, enhancing tool calling, long-chain execution, and scheduled tasks. It ranks first among domestic models on the proprietary ZClawBench benchmark and comes with enterprise security management and promotional packages.

Full Report

Contents7 sections

We are introducing GLM-5-Turbo—a foundation model deeply optimized for OpenClaw "Lobster" scenarios. Users who have experienced OpenClaw share a common sentiment: the model can chat well, but doesn't necessarily work well. The root cause lies not in the framework, but in the underlying model itself. General-purpose large models tend to lose momentum in long-chain tasks when entering real-world, complex Agent scenarios. No amount of harnessing can compensate for a stronger base model. From the training phase, GLM-5-Turbo is specifically optimized for the core requirements of Lobster tasks, enhancing key capabilities such as tool calling, instruction following, scheduled and persistent tasks, and long-chain execution. It achieved first place among domestic models on our proprietary benchmark, ZClawBench; multiple major internet companies gave it high praise during internal testing in Lobster scenarios. We are also launching the Claw Lobster Package and the Claw Security Management System for enterprise-grade scenarios.

Lobster-Native Model

Lobster tasks are not simple Q&A interactions; they typically involve multi-turn understanding, task decomposition, tool calling, state continuity, time-triggered actions, and persistent execution within long-chain workflows. Even if general-purpose models perform excellently in conversational abilities, they often suffer from issues like instruction-following deviations, unstable tool calls, and mid-task stalls once deployed in real Lobster scenarios. To fundamentally solve these challenges, deep optimization at the foundation model layer is essential—this is precisely the starting point for GLM-5-Turbo. From training data construction to optimization objective design, we build realistic Agent workflows to ensure the model's executability in Lobster tasks. We focus on enhancing the following core capabilities:

  • Tool Calling – Reliable invocation without failures. Strengthening the ability to call external tools and various Skills.
  • Instruction Following – Better decomposition of complex instructions. More accurate understanding and breakdown of complex, multi-layered, long-chain instructions, supporting goal identification, step planning, and multi-agent collaboration.
  • Scheduled & Persistent Tasks – Better temporal awareness, no interruptions for long tasks. Key optimizations for scenarios involving timed triggers and long-running processes, enabling better comprehension of time-dimensional requirements.
  • High-Throughput Long-Chain Execution – More stable performance. Further improving execution efficiency and stability for tasks with high data throughput and long chains, suitable for extended business processes.

Lobster Scenario Benchmark: ZClawBench

As OpenClaw gains popularity, evaluating model capabilities in Lobster scenarios has become an industry-wide focus. Based on the analysis of numerous real-world use cases from OpenClaw, we release ZClawBench, an end-to-end Agent evaluation benchmark for Lobster scenarios. Currently, OpenClaw task types cover diverse areas including installation configuration, code development, information gathering, data analysis, and content creation. The user base has expanded from early developers to include productivity-focused professionals, financial practitioners, DevOps engineers, content creators, and research analysts. Meanwhile, the usage rate of Skills has rapidly grown from 26% to 45% in a short period, indicating that Agent capabilities are evolving toward a modular and skill-based ecosystem. Evaluation results based on this benchmark show that GLM-5-Turbo demonstrates significant improvements over GLM-5 in OpenClaw scenarios, leading across multiple key tasks compared to several mainstream models.

The question bank and test trajectories for ZClawBench are fully open-sourced, inviting the industry to jointly verify and refine them. Real user experience is the ultimate touchstone. Under the codename Pony-Alpha-2, GLM-5-Turbo was integrated into AutoClaw (AoLong). In blind user tests, 90% of respondents rated GLM-5-Turbo's performance higher than other domestic models. During the beta phase, major internet companies within the Lobster ecosystem gave GLM-5-Turbo high praise:

"Precise tool calling combined with reliable long-range task execution perfectly drives cross-application data flow and system API control on desktop-level Agents. Excellent performance in tackling the 'hallucination' pain point in office scenarios!" — Alibaba QoderWork Model Evaluation Team

"Unmatched instruction following, more precise and reliable handling of long tasks, significantly boosting the success rate of multi-agent construction." — Coze Evaluation Team

"Stable performance in scheduled and long-chain tasks, demonstrating clear first-tier advantages compared to similar models. Very reliable." — Meituan Team

"Extremely fast high-throughput execution without stalling, leading peers in response stability, truly solving Agent execution challenges." — Kuaishou Wanqing Evaluation Team

In programming scenarios, GLM-5-Turbo continues the Agentic Engineering performance legacy of the GLM series:

"Programming capabilities remain firmly in the first tier, outperforming similar models in Coding Agent tests, with very robust handling of long-chain code tasks." — ByteDance TRAE Team

GLM-5-Turbo has been integrated into the Mechanical Revolution Box under iSoftStone, marking the global debut of the Mechanical Revolution "Lobster Box" powered by the GLM model, creating a native AI Agent terminal experience.

Lobster Packages to Fuel Your AI Colleagues

Lobsters are highly productive workers, which naturally means they consume tokens. A typical Lobster task, such as organizing cross-departmental meeting minutes and distributing action items, may involve dozens of rounds of tool calls and context continuity, consuming far more tokens than standard conversations. For enterprises and individuals using Lobsters, token costs have become the biggest bottleneck. To enable "token freedom" for Lobster users, Zhipu is launching Lobster packages based on GLM-5-Turbo, available in Personal and Team editions. Whether you are a solo entrepreneur, a startup team, or a large-to-medium enterprise, you can flexibly subscribe to packages according to your business scale.

Each account is limited to a maximum of 5 purchases. Enterprises can flexibly subscribe to the Lobster Plan Team Edition based on their actual headcount, ensuring sufficient token allocation and stable AI employee availability at a controlled cost. Furthermore, building upon the Lobster Plan, we have specifically developed an enterprise-grade security management system for OpenClaw called "Claw for Enterprise Security." This system supports unified scheduling and permission orchestration for various types of Claws—whether for information gathering, data analysis, code development, or cross-system operations. Enterprises can flexibly combine Claw capabilities according to their business processes and monitor the execution paths, tool invocation chains, and resource consumption of every Agent task in real-time through a centralized control console, making complex multi-Agent collaboration no longer a black box.

In terms of security assurance, the system provides multi-layered protection mechanisms: role-based access control ensures that each Claw can only invoke tools and access data within its authorized scope; all operations generate complete audit logs to meet corporate compliance requirements; sensitive business data is encrypted throughout the entire process, strictly adhering to data localization requirements and never leaving the enterprise's controllable boundaries; manual approval mechanisms are retained at critical business nodes to prevent uncontrollable business risks arising from automated workflows. The transformation underway is that Agents represented by OpenClaw are entering organizations, providing every enterprise with 24/7 online, on-demand AI employees. Simultaneously, they serve as personal AI assistants for everyone, truly freeing people from procedural and repetitive trivialities. We are witnessing a key leap where large models evolve from efficiency tools into enterprise workforce components.

Access and Usage

GLM-5-Turbo is now live. Developers and enterprise users can directly call the API via Zhipu's open platform BigModel.cn or access it through the Lobster Plan. GLM Coding Plan Max currently supports GLM-5-Turbo; the Pro plan will support it as soon as possible this month, and the Lite plan will support it in April. Coding Plan can also be used for OpenClaw trial experiences; however, for long-term stable operation, continuous background execution, or higher-load Agent scenarios, the Lobster Plan is recommended. As advanced models benchmarked against Claude Opus, GLM-5 and GLM-5-Turbo consume quota at a coefficient of "3x during peak hours, 2x during off-peak hours." As a limited-time benefit, GLM-5-Turbo will only deduct 1x quota during off-peak hours until the end of April. AutoClaw (Ao Long) defaults to GLM-5-Turbo, allowing non-programmers to experience GLM-5-Turbo's Lobster capabilities with a certain amount of free quota. From today until March 22, Ao Long offers a limited-time promotion where top-up packs enjoy up to a 66% discount (top-up packs are valid for 7 days from purchase).

Official API Access

Zhipu Open Platform: BigModel.cn
Z.ai: api.z.ai

GLM Coding Plan

Coding Plan Max has included GLM-5-Turbo.
The Pro version will support it as soon as possible this month.

Online Experience

AutoClaw (Ao Long) Client: https://autoglm.zhipuai.cn/autoclaw
Z.ai: z.ai
Zhipu Qingyan APP/Web Version