GLM-5.2, Day-0 on FriendliAI Model APIs: The Strongest Open Model for Long-Horizon Agentic and Coding Tasks

GLM-5.2 is Z.ai's most capable open-weight model to date, and it is built for the workloads that define modern AI: autonomous coding agents, multi-step tool use, and long-horizon tasks that have to stay coherent over hours, not seconds. It inherits the 744B-parameter Mixture-of-Experts foundation of the GLM-5 series — activating roughly 40B parameters per token — and pairs it with a genuinely usable 1M-token context window and a new effort-level system that lets you dial reasoning depth up or down per request.

FriendliAI is proud to launch Day-0 support for GLM-5.2 via Model APIs. OpenAI-compatible, production-grade, and tuned for the MoE and long-context patterns that make this model efficient. No capacity planning, no GPU orchestration, no cold starts to babysit. Point your agent at FriendliAI and go.

Whether you're building a coding agent that refactors an entire repository in a single session, an orchestration layer that fans out across dozens of tools, or a long-running research agent, GLM-5.2 on FriendliAI gives you frontier-class open intelligence at a fraction of closed-model cost.

How GLM-5.2 performs

Across a broad suite of coding, repository, and agentic tool-use benchmarks — all evaluated under maximum thinking effort, GLM-5.2 establishes itself as the leading open-weight model and trades blows with frontier closed systems.

Image 1) Z.ai GLM-5.2 technical blog: z.ai/blog/glm-5.2

A few results stand out for agent builders:

• Terminal-Bench 2.1: 81.0 — GLM-5.2 is the first open-weight model to cross the 80% mark, landing within a few points of the best closed models and far ahead of every other open option.
• MCP-Atlas: 77.0 — essentially tied with the top frontier model on tool orchestration over the Model Context Protocol, and ahead of GPT-5.5 and Gemini 3.1 Pro.
• SWE-bench Pro: 62.1 and ProgramBench: 63.7 — leading the open field and ahead of GPT-5.5 and Gemini 3.1 Pro on real-world software engineering and program synthesis.
• Humanity's Last Exam (with tools): 54.7 — competitive with frontier closed models on hard, tool-augmented reasoning.

Where GLM-5.2 really shows its character, though, is in how it scales with effort. Reasoning effort buys accuracy, and GLM-5.2 converts additional output tokens into real gains.

Image 2) Z.ai GLM-5.2 technical blog: z.ai/blog/glm-5.2

Averaged over Terminal-Bench 2.1, DeepSWE, and SWE-Atlas QnA, GLM-5.2 climbs from its Non-Thinking baseline to roughly 75% at Max effort, a steep improvement over GLM-5.1 and competitive with much larger closed models on the same effort curve. The takeaway for production: you control the trade-off. Run High effort for fast, economical agent loops, and reach for Max only on the hard problems that justify it — exactly the kind of cost-versus-quality tuning that matters when you're running thousands of agent sessions a day.

What you can build with GLM-5.2 on FriendliAI

GLM-5.2 is purpose-built for agents that operate over long horizons and large contexts. FriendliAI's high-performance Model API unlocks these capabilities at production scale, pay-per-token.

Autonomous coding agents

Point Claude Code, Kilo, Cursor, Cline, or any OpenAI-compatible coding agent at GLM-5.2 on FriendliAI and let it run. With Max effort and a 1M-token context, the agent can reason over an entire codebase, track cross-file dependencies, and carry a task through implementation, testing, and iteration in a single coherent session.

Repository-scale refactors

Load a mid-sized repo into one context window and let GLM-5.2 reason across it holistically — migrations, framework upgrades, large-scale refactors — without the accuracy loss that comes from chunking a project across many short calls.

Long-horizon, multi-tool agents

GLM-5.2's strength on MCP-Atlas and Tool-Decathlon makes it a strong engine for orchestration agents that fan out across many tools and APIs over long sessions, maintaining state and making sound decisions step after step.

Cost-sensitive agentic pipelines at scale

When you're running thousands of agent sessions a day, token economics decide what's viable. GLM-5.2's open-model pricing on FriendliAI — combined with effort-level control — lets you ship agentic features that would be cost-prohibitive on closed frontier APIs.

Why run GLM-5.2 on FriendliAI Model APIs

A frontier open model is only as good as the infrastructure serving it. GLM-5.2's MoE architecture, sparse-attention long-context path, and long agentic sessions all demand an inference stack engineered for exactly these patterns — which is what FriendliAI is built for.

With FriendliAI Model APIs, you get:

• Serverless, pay-per-token access — no GPU provisioning, no capacity planning, no idle cost. Call GLM-5.2 like any API and pay only for the tokens you use.
• 2–5× faster output token speed — FriendliAI's inference engine is tuned for high-throughput MoE serving and efficient long-context decoding, so your agents finish sooner.
• 50–90% lower inference cost — more tokens per GPU means a better tokens-per-dollar number, on top of GLM-5.2's already open-model economics.
• 99.99% uptime SLA — production-grade reliability for agent workloads you can't afford to have stall.
• OpenAI-compatible API — drop GLM-5.2 into your existing agent stack with a model-name change.

Get started with GLM-5.2 on FriendliAI Model APIs

GLM-5.2 is available now through FriendliAI Model APIs. You can plug it straight into your coding agent or call it directly from your own code.

Prerequisites

• A FriendliAI account
• A FriendliAI API key (create one in the Friendli Suite dashboard)

Use GLM-5.2 in your coding agent

GLM-5.2 works out of the box with Claude Code, Cline, Cursor, Kilo Code, OpenClaw, OpenCode, and Hermes — any agent that can point at a custom endpoint. Connect your agent to FriendliAI's Model API base URL, https://api.friendli.ai/serverless, and set the model to zai-org/GLM-5.2.

For example, to use GLM-5.2 in Claude Code, open your ~/.claude/settings.json file and add:

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "<YOUR_FRIENDLIAI_API_KEY>",
    "ANTHROPIC_BASE_URL": "https://api.friendli.ai/serverless"
  },
  "model": "zai-org/GLM-5.2"
}

Save the file, open Claude Code, and send a prompt — your requests now run on GLM-5.2 via FriendliAI. For per-agent setup, see FriendliAI Docs › Use Your Agent with FriendliAI.

Take open frontier intelligence to production with FriendliAI

GLM-5.2 represents a new high-water mark for open-weight models — frontier-class coding and agentic reasoning, a usable 1M-token context, and an MIT license that puts you fully in control. FriendliAI is the platform to take it to production: fast, cost-efficient, reliable inference, available the moment the model launches.

👉 Start building with GLM-5.2 on FriendliAI today — call it through Model APIs in minutes!