This has been a long-standing problem. Many models are getting truncated context, or in other words, the nominal context of say 200k, 250k or 275k tokens, is larger than the context that actually seems to be available for use.
This model, Sonnet 4.5, should have a 200k context window, but it has for quite some time now, shown only 175k. The issue is even worse with GPT-5 Codex, which is supposed to have a 275k context window, but only gets 200k:
This does seem to affect when context summarization happens. I rarely get beyond about 87% context usage anyway, maybe 90%, before context summarization happens. That means we are first not able to use all the available context before summarization occurs, and for many chats summarization may never need to occur, but also it results in summarization occurring more frequently, resulting in longer wait times while the agent summarizes context.
Steps to Reproduce
Select model.
View context usage.
Expected Behavior
Available context should match the nominal context available for each model.
Summarization should not be triggered prematurely based on an incorrect context usage meter.
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
The reduced available context window may be due to reserved tokens for:
space for response generation
space for reasoning
system prompts and instructions
safety buffers
However, the difference you noted (25–75k tokens) seems significant, especially if it causes premature summarization.
I’ll pass this to the team to determine whether the reduced available context is intentional or a bug. The team will check whether the context window calculations work correctly, or if there’s an issue with how the available context is calculated or displayed.
I think, if you guys are reserving some amount of context for those purposes, that reservation should be made evident in this control. A different color, perhaps, in the radial indicator to show how much of the usage is for reservations. Then in the popup, report out of the total context, how much is reserve vs. user usage. IF we have a 275k context window, and 10ktok are reserved, and we then have 265ktok for our usage, that needs to be made clearer.
I do think that if you have a 275ktok window, but are losing 75k for some reason (reservation or whatever), that is a lot of context to be losing. I also can’t imagine its all reserve, so hopefully its just a calculation glitch. It does seem, though, that context gets summarized around 90% usage on that indicator. Sometimes it edges higher before summarization, but, if there is a buffer that basically forces summarization early, AND there are also reserves…we are kind of losing a lot of the context window here…
