GLM-5: Features, Pricing, and Agentic Coding
GLM-5 is an open-weight AI language model from Z.ai designed for systems engineering, long-horizon agents, and coding work. It uses a mixture-of-experts architecture with 744 billion total parameters and 40 billion active parameters, plus a 200K-token context window. The weights are published under MIT, while hosted API use has a separate price and service layer.
The family now includes later models. That makes exact version labeling important when you compare benchmark claims, serving costs, or a Verdent configuration.
What Is GLM-5
It combines mixture-of-experts routing with DeepSeek Sparse Attention. The published API identifier is glm-5, and the provider lists up to 128K output tokens within its context limits. The architecture aims to allocate a smaller active parameter set per token while retaining a much larger total capacity.
The current standard API table lists $1 per million input tokens, $0.20 per million cached input tokens, and $3.20 per million output tokens. At those rates, cache reuse can materially affect a long agent task.
SWE-bench Performance
The published evaluation reports a 77.8 score on SWE-bench Verified. This is a vendor-reported result tied to a specific setup, not a promise for every repository. Your outcome will depend on task selection, scaffolding, tools, retries, and review rules.
Use a small evaluation matrix:
| Test | Acceptance signal |
|---|---|
| Bug fix | Root cause found and regression test added |
| Refactor | Public behavior preserved and diff stays bounded |
| Tool loop | Commands are relevant and failures are recovered |
| Long task | Plan remains consistent after multiple tool calls |
GLM-5 vs DeepSeek V3.2 vs Kimi K2.5
All three models have published weights, but they emphasize different workflows. This model focuses on systems engineering and long-horizon agents. DeepSeek V3.2 combines thinking modes with tool use. Kimi K2.5 adds native multimodal input and visual coding capabilities.
Compare the exact checkpoint, license, context implementation, and serving stack. Do not transfer a score from one harness to another or assume a larger context automatically produces better code changes.
MIT License
The published weights use the MIT license. That supports broad use, modification, and redistribution while requiring preservation of the license notice. A hosted endpoint can still impose separate billing, availability, retention, and acceptable-use terms.
For self-managed deployment, document the weight source, quantization, inference framework, and any modifications. The license makes deployment flexible, but it does not remove infrastructure or governance work.
Agentic Coding Use Cases
The model is most relevant when the task requires a plan, repeated tool calls, and state carried across a longer execution. Practical evaluation prompts include:
- “Map this service dependency, identify the failing boundary, and propose a repair plan before editing.”
- “Upgrade this library across the monorepo, run the relevant tests, and summarize unresolved incompatibilities.”
- “Implement the approved API change, keep backward compatibility, and show evidence for every acceptance criterion.”
Watch for drift after long tool sequences, unnecessary file changes, and weak recovery from failed commands. Review effort is as important as task completion.
Using GLM-5 in Verdent
Verdent added support for this version in February 2026, while today's built-in catalog lists later GLM models. Teams that specifically need the earlier release can use the documented Z.ai BYOA path, subject to provider access.
Start with one tool-use flow before assigning a long repository change.
Stress-Test Tool Recovery
Connect the Z.ai provider, assign one bounded coding task, and introduce a failed command. Compare the recovery plan, changed files, and test evidence.
Frequently Asked Questions
Are the model weights available for deployment?
Yes. The weights are published, but deployment requirements depend on the precision, serving framework, hardware layout, and throughput target.
What is the documented context window?
The model documents a 200K-token context window, with up to 128K output tokens. A host may apply lower operational limits.
What is the API model name?
The documented API model name is glm-5, used with the Z.ai endpoint and provider credentials.
Does the model require self-hosting?
No. Teams can use a hosted API or deploy published weights. The right path depends on operations capacity, data controls, latency, and cost.