@0xHACKS I cannot be much clearer. With MAX mode disabled, we will cap the maximum context that is sent to the LLM to ensure you aren’t accidentally using all 1 million tokens of Gemini (as an example) without meaning too.
With MAX mode enabled you get the model as-is, with its full context window.
Besides this specific feature (which has been in place since Cursor added agent mode), and our custom system prompt, the model is exactly how it would be from the API provider. We do nothing to specifically cut it’s intelligence, input or output token limits beyond as described.
Short of showing you the code thats responsible for this, this is as definitive as I think I can be
This has been the case for months, this is not new. I am unhappy with Cursor as much as the next person (these days) but the thing you are claiming to “catch him admitting” has been common-knowledge for every user. It says it clearly in the model-selection menu in the Cursor app. MAX mode is billed at API costs because it includes the full context window!!!
They have never changed this other than the early-early days of Cursor.
When I started using claude code this month I was shocked by the accuracy of the model as opposed to cursor’s equivalent.
Using memory MCP, one hint is my memory MCP states models having an explcit rule to limit to 3 the amount of changes by file, crippling on purpose the abilities of their models to change meaningfully several files, and rather burn asap the 25 actions for the next “resume the conversation”.
@condor I’m using Opus since yesterday for coding outside of Cursor and haven’t seen any issues. but there could a clash between Cursor agents and Opus internal functionality
You’re framing it like I just discovered MAX mode, I didn’t. I’ve been here since the beginning. My point is simple : when full model behavior is gated behind MAX, it is modified by default, and that deserves clarity, not spin. Saying “it’s always been like this” doesn’t make it any less valid to call out now.
Also… who came up with this gibberish username? “2jfs904judsw20600jik”, definitely smells like an employee sock puppet account.
Could you post a full separate bug report on your last Opus 4 issues with a privacy disabled Request ID so we can have a look in detail?
Without specifics to look into its hard to tell what issues occur or why as its not affecting every user.
Overall I recommend not to engage a model in back and forth if it starts going off the rails. Often creating a new chat and guiding the model to prevent issue from reoccurring helps, but again it depends on how you use it.
(Note that there is no @Cursor user which is why its greyed out, you can tag me as @condor )
