GLM 5.2 on MAX context, does not really have 1M context size

[Bart_van_Berkel]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

When selecting glm-5.2 with MAX context (1M) the chat summarizes around 200K context. This is not a BYOK model. It is included in our company subscription. I don’t care that 1M context size is not possible, this could be because the model creator is not really honest about max context size, or it could be because cursor cannot run the model with this size (or some other reason). That I do not care about, what this issue is about, is the expectation that summary will not happen for a long time, when you have about 150K context, and the total size is supposed to be 1M. This expectation is then not met, when chat summary happens when passing the 200K threshold. This keeps happening if you keep trying to load in far more context (which you could do, based on your 1M context size expectation).
So either the real size should be shown, so that my expectation is not 1M. Or something should be fixed, so that the context size is indeed around 1M and we don’t summarize around 200K.

Steps to Reproduce

Open a cursor agent with glm 5.2 MAX context, hosted by cursor itself. Then ask the agent to load in random filler. See that the context grows, but never further than ~200K. Chat summary will happen.

Another reproduction path is to open a fable 1M context agent, load up to around 300K context (for instance random filler), and then switch to glm-5.2 max context, and ask it a simple question. Chat summary will happen immediately.

Expected Behavior

I would expect that either the correct context size is shown (maybe 1M context is simply not actually possible with GLM 5.2), or that chat summary does not happens around 200K context size, but more around 1M.

Screenshots / Screen Recordings

Operating System

MacOS

Version Information

Version: 3.11.25 (Universal)
VS Code Extension API: 1.125.0
Commit: fc2563ec93d793fc275eef734405a4fdf8b47b20
Date: 2026-07-15T01:27:41.754Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Darwin arm64 25.5.0

For AI issues: which model did you use?

GLM 5.2 (subject of issue), Fable 5 (used to show issue)

For AI issues: add Request ID with privacy disabled

9732e07c-8c92-439a-964b-621107fe4585

Does this stop you from using Cursor

No - Cursor works, but with this issue

[Colin]

GLM 5.2 on MAX context, does not really have 1M context size

Ah, you’re right. I was working with some outdated information. GLM 5.2 does have access to a 1M context window, but there appears to be a bug (exactly as you described) that kicks off summarization at 200k. We have a bug filed for this.

Wrong answer, see below.

Hi @Bart_van_Berkel!

GLM 5.2 only has one context window size in Cursor, 200k! Can you share where you’re selecting a larger context window?

Models that have adjustable context windows are showin in the model selector.

For GLM 5.2 you can only adjust the reasoning effort.

When switching between models of various context windows, the Context Usage report will not update until a turn on the new model is completed. This is why you see a larger context window in teh report when switching between Fable and GLM 5.2.

That being said, maybe we could do something better here to make it clear that the existing context is larger than the model you’re switching to!

[Bart_van_Berkel]

Hi @colin. Thank you for the reply.
I select the MAX mode with the toggle, when selecting the model. So not the max mode when it comes to ‘thinking effort’. But the toggle for context size. It then shows in the context overview that there should be a 1M context.
It shows this 1M total size, not just the first turn. But on every turn. It just summarizes when it reaches 200K. But still shows 1M total.

[Colin]

Ah, you’re right. I was working with some outdated information.

GLM 5.2 should have access to a 1M context window, but there appears to be a bug (exactly as you described) that kicks off summarization at 200k.

We have a bug filed for this.