GLM 4.7 Max sub is slow and dumb

[moonray]

Description

The free version was actually pretty good for tool calling and orchestration. Now that I'm using the z.ai max sub, it feels slow and dumb. That has to be a bug.

Is there some config that was applied to the free model by default, such as thinking levels (don't see any high thinking variants in opencode) or anything else? Using GLM 4.7 max sub feels like significant downgrade to the free version.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Differences between the free and paid versions of GLM 4.7 #6074: Differences between the free and paid versions of GLM 4.7

Feel free to ignore if this doesn't address your specific case.

[moonray]

I'm wondering if it's related to this:

From docs.z.ai

Preserved Thinking

This capability is enabled by default on the Coding Plan endpoint and disabled by default on the standard API endpoint. If you want to enable Preserved Thinking in your product (primarily recommended for coding/agent scenarios), you can turn it on for the API endpoint by setting “clear_thinking”: false, and you must return the complete, unmodified reasoning_content back to the API.
All consecutive reasoning_content blocks must exactly match the original sequence generated by the model during the initial request. Do not reorder or edit these blocks; otherwise, performance may degrade and cache hit rates may be affected.

Turn-level Thinking

“Turn-level Thinking” is a capability that lets you control reasoning computation on a per-turn basis: within the same session, each request can independently choose to enable or disable thinking. This is a new capability introduced in GLM-4.7, with the following advantages:

• More flexible cost/latency control: For lightweight turns like “asking a fact” or “tweaking wording,” you can disable thinking to get faster responses; for heavier tasks like “complex planning,” “multi-constraint reasoning,” or “code debugging,” you can enable thinking to improve accuracy and stability.
• Smoother multi-turn experience: The thinking switch can be toggled at any point within a session. The model stays coherent across turns and keeps a consistent output style, making it feel “smarter when things are hard, faster when things are simple.”
• Better for agent/tool-use scenarios: On turns that require quick tool execution, you can reduce reasoning overhead; on turns that require making decisions based on tool results, you can turn on deeper thinking—dynamically balancing efficiency and quality.

[vladlearns]

Same here. Very slow and very dumb

[ravshan-oxagile]

Maybe this is because openai api is used rather than anthropic api. Also afaik z.ai supports only chat completions api (the legacy one) not the new responses api

[maarteNNNN]

I subscribed for the Max for one year after a worse than Claude performance but better pricing using GLM 4.6. Since 4.7 it's just obnoxious how bad this is performing. It's slow as f*** and not performing nearly as close as it used to be. I feel cheated.

On contacting Z.ai when I received some errors about rate-limiting they replied with the following:

However, we notice that you receive error codes in Chinese, which is un-expected. The api endpoint you are using 'https://bigmodel.cn/claude-code' is incorrect, since you are a user in z.ai platform, please ensure to use the correct api endpoint with domain name api.z.ai.
For Claude Code: https://api.z.ai/api/anthropic
For Other Tools: https://api.z.ai/api/coding/paas/v4

[maarteNNNN]

I seemed to be able to increase performance with manually putting it in the config:

    "zhipuai-coding-plan": {
      "options": {
        "baseURL": "https://api.z.ai/api/coding/paas/v4"
      },
      "models": {
        "glm-4.7": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5-air": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5-flash": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5v": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.6": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.6v": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.6v-flash": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        }
      }
    },
    "zai-coding-plan": {
      "options": {
        "baseURL": "https://api.z.ai/api/coding/paas/v4"
      },
      "models": {
        "glm-4.7": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.7-flash": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.6": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.6v": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5-air": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5-flash": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        },
        "glm-4.5v": {
          "limit": {
            "context": 100000,
            "output": 4096
          }
        }
      }
    }
  },

I reduced the context as well for faster speed.

Would it by default be connecting to Chinese servers?

I also don't know what's the difference between zai-coding-plan and zhipuai-coding-plan...

When removing my API key it really seems to be a Chinese endpoint as per error message: 您的GLM Coding Plan套餐已到期，暂无法使用，前往官方续订后即 可恢复https://bigmodel.cn/claude-code。

I will do some more testing hopefully it's resolved.

[TotalLag]

Max plan user as well. The config worked for me too and opencode is zippy as well. I know on GLM docs they make mention to use the specific URL so I thought it was default on opencode. Thank you!

What also helped me was adding in concurrency settings for all the models so opencode can make use of its queues.

Zhipu is the China branch of zai.

[diegohb]

Right, is this not accounted for in OpenCode fault auth flow for z?

Instructions on Z website says default opencode auth flow works: https://docs.z.ai/devpack/tool/opencode

[mikaelj]

@TotalLag could you please share the settings you use? specifically the part about concurrency.

[maarteNNNN]

as per config:

    "zai-coding-plan": {
      "options": {
        "baseURL": "https://api.z.ai/api/coding/paas/v4",
        "concurrency": 5
      },

@TotalLag what number did you define and are you getting good results? Mine still aren't that great compared to Claude Sonnet using GH Copilot...

[moonray]

I checked my outgoing network traffic, and my endpoint is not pointing to the Chinese servers, but to the correct international servers (I didn't add the z.ai models to my config).
So that's definitely not the issue for me.