In the cited Z.ai documentation, the API model identifier is glm-4.6; “Thinking” is a setting, not a separate model ID. Z.ai’s migration example enables it with thinking={"type": "enabled"}. The documentation also describes enabled thinking as dynamic: GLM-4.6 decides whether a prompt calls for it.
Enable thinking in a GLM-4.6 API request
Z.ai’s GLM-4.6 migration guide shows this SDK pattern:
resp = client.chat.completions.create(
model="glm-4.6",
messages=[
{"role": "user", "content": "Design a three-tier microservice architecture for me"}
],
thinking={"type": "enabled"}
)
Use a Z.ai client configured with your API key. This example supports the model ID and enabled-thinking setting for the SDK. The materials cited here do not establish a GLM-4.6-specific REST request, bearer-authentication recipe, disabled value, or streaming response fields, so those details are not included as implementation instructions.
GLM-4.6 is not GLM-4.6V
GLM-4.6V is a distinct multimodal model identifier. Its documentation also includes a thinking setting, but that example is for glm-4.6v and should not be substituted for the GLM-4.6 SDK call above.
Is GLM-4.6 worth evaluating?
Consider testing GLM-4.6 if your workload involves multi-step reasoning, coding, long-context text, or tool-using agents. Z.ai’s model documentation lists a maximum context of 200K tokens and a maximum output of 128K tokens; check the limits that apply to your specific API use before planning around them.
Treat benchmark results as a reason to run your own evaluation, not a performance guarantee. In its release post, Z.ai reports a 48.6% win rate against Claude Sonnet 4 on its human-evaluated, Docker-isolated CC-Bench tasks, and about 15% fewer tokens than GLM-4.5 on those task trajectories. These are vendor-reported results from described evaluations; they do not guarantee lower cost, faster responses, or better results on your workload.
A practical selection test:
- Choose representative tasks: Use examples from your actual workload, such as debugging tickets, long-document questions, or tasks that require tool calls.
- Compare thinking settings: Test enabled thinking on the same prompts you use for your baseline. Record task success, unsupported answers, completion time, and token use.
- Evaluate tool execution separately: If your application calls functions or external services, measure the whole workflow—not only the model’s written response.
- Check operational fit: Confirm availability, limits, latency, and billing for the API plan you intend to use. The cited materials do not establish pricing for your account or usage channel.
Bottom line: In the cited Z.ai API instructions, use glm-4.6 as the model ID and enable thinking with the documented setting. GLM-4.6 may be worth evaluating for complex reasoning and coding, but make the choice using representative task results and verified production constraints.
Sources
Migrate to GLM-4.6 — Z.AI Developer Documentation
Deep Thinking — Z.AI Developer Documentation
GLM-4.6 Overview — Z.AI Developer Documentation
GLM-4.6: Advanced Agentic, Reasoning and Coding Capabilities — Z.ai