GLM-5.2 is now available on CloudCode.ONE
GLM-5.2 is now available on CloudCode.ONE
July 2026
Zhipu's new flagship coding model GLM-5.2 just landed in the CloudCode.ONE model selection. It's the strongest open-weights coding model available right now, and you can use it today through the same Anthropic- and OpenAI-compatible endpoints you're already pointing Claude Code, Cline, Roo Code, or Harness at.
Why GLM-5.2 is a big deal
A 1M-token context window that actually holds up. Plenty of models advertise long context; GLM-5.2 was specifically trained for long-horizon coding agent work, so it keeps module boundaries, API contracts, and earlier architectural decisions intact deep into a session instead of gradually forgetting them. That's exactly the failure mode that kills multi-hour agentic runs.
Frontier-adjacent benchmark results. GLM-5.2 scores 81.0 on Terminal-Bench 2.1 — within a few points of Claude Opus 4.8's 85.0 and ahead of Gemini 3.1 Pro — and 62.1 on SWE-bench Pro, a large jump over GLM-5.1. On long-horizon suites like FrontierSWE it trails Opus 4.8 by roughly a single percent while beating GPT-5.5 and Opus 4.7 on several benchmarks. For an open-weights model at open-weights pricing, that's a remarkable place to be.
Built for real engineering discipline. Early users report it's noticeably better at sticking to hard constraints — lint rules, build commands, "don't touch this API contract", "don't commit" — across long multi-turn runs. If you drive it from a CLAUDE.md or Agent.md with your repo conventions, it holds the line instead of drifting out of scope.
Thinking mode, tool calling, structured output, and context caching are all supported, so it slots straight into agentic loops.
What it's great at on CloudCode.ONE
- Whole-project takeover: dump an entire codebase — backend, frontend, configs, tests, docs — into context and have it produce an architecture map and refactoring plan in one pass
- Long-horizon refactors: module decoupling, API migrations, cross-language rewrites that need a plan, staged execution, and verification
- Mobile dev loops: it can go from Kotlin implementation to real-device debugging with ADB, logcat, and screenshots
- Research reproduction: turning a paper plus dataset into a runnable PyTorch project
How to use it
Select glm-5.2 in your CloudCode.ONE dashboard, point your client at api.cloudcode.one as usual, and go. It works with:
- Harness, our free desktop coding agent for Windows and Ubuntu
- Claude Code (
ANTHROPIC_BASE_URL=https://api.cloudcode.one)
- Cline, Roo Code, OpenCode, and anything else that speaks the OpenAI or Anthropic API
One note: GLM-5.2 is text-only. On CloudCode.ONE that doesn't slow you down — paste a screenshot, UI mockup, or architecture diagram into your prompt and our automatic vision routing hands the image to a vision-capable model, then keeps your coding session on GLM-5.2. You get flagship open-weights coding and image understanding in one workflow, which is something you won't get pointing your tools at the model directly.
As always: pay-as-you-go, top up from $2, no subscription, credit never expires.
Create an account and try it — your API key is provisioned instantly.
CloudCode.ONE — pay-as-you-go agentic coding through Anthropic- and OpenAI-compatible APIs.