I mean, LLMs still aren’t quite capable of building large apps from just a few vague prompts. That said, Claude 3.5 Sonnet (20241022) seemed pretty effective at the time—either because it handled my tasks well or was generally more capable. However, as I continued using 3.7 Sonnet, I noticed it often breaks existing features when adding new ones. I’ve also seen many others complain about the poor performance of these models. I just want an official statement from Cursor on whether the models are quantized or downgraded in any way to maximize profit—especially since Anthropic’s API pricing is quite high compared to others, even though Cursor might receive a discount due to its usage volume.
25 Likes
I am also very curious about this question
4 Likes
Yeah IMO unless they start to offer the option of running models locally(and yes I know you can with Ollama and external tools) then I think cursor is just going to get worse, in terms of offering more money grabbing features while lessening the value/usefullness of existing ones.
4 Likes
I don’t think they quantized or downgraded any models, but it’s very likely they removed significant amount of context while sending requests to Anthropic to save money (for non-MAX usage). When LLM loses context it does not know the “global picture” so it often makes stupid mistakes or “fix” code which used to be working properly.
I think Cursor team needs to communicate better with users on what’s going on here. Claude 3.7 is expensive which we all know. Paying some extra $? fine. But we just don’t want to use something that can’t be controlled.
Also, it doesn’t make any sense to charge gemini-2.5 MAX the same as Claude 3.7 MAX. Google makes it free at the moment, doesn’t it?
14 Likes
Gemini 2.5 pro is experimental, which itself doesn’t justify the $0.05 request/tool-calling, correct? That’s the reason why Google makes it free for now. It’s free for everyone, including users with paid plan, so I don’t think Cursor is paying anything to Google. Regarding the agent calls and functionalities that Curson builds upon it, yes we respect these efforts but didn’t premium users already pay for the 500 monthly fast requests?
7 Likes
Ok, Gemini just announced the pricing 15 hours ago (still free yesterday):
However it’s still generally cheaper than Claude 3.7
I am not saying Cursor is bad. It is great. I am just confused by their pricing scheme. On cursor’s pricing page, under $20, $40 plan, did it mention that you can’t use Claude 3.7 or Gemini 2.5 unless you pay additional 5 cents for every tool calling?
5 Likes
whether they are or aren’t, it’s not a good look.
The default 3.7 in Cursor is noticeably stupider than Desktop Claude with file-edit mcp. In fact it’s been more reliable at building features and fixing complex bugs than with Cursor’s token-optimizing-thumb-on-the-scale-back-end …
If the vapid useless automated “composer” summaries are any indication – Cursor’s lobotomizing Claude with their go-between models.
6 Likes
The only difference between MAX and non-MAX is the context window size, at least what they tell us.
3 Likes
I, too, have noticed degrading performance when using Cursor. Recently, the models tend to fail completely at the given tasks, when I’m using Claude 3.7 Sonnet, unless I’m using the ‘Max’ mode – even if I’m using ‘thinking’ mode. I remember having less unsuccessful responses with 3.5 Sonnet.
However, I can’t say for certain that Cursor are nerfing the models. It could simply be that my project is too large and/or complex for the smaller context window. Hopefully, we can get a response to this issue from the Cursor team.
5 Likes
Why does MAX mode even exist? 
Isn’t MAX mode basically just the model’s default performance to begin with, rather than something genuinely special?
It seems like the so-called “Default” mode is actually a limited-context version—not the standard one.
Please be transparent with your users.
12 Likes
MAX mode provides enhanced capabilities beyond default model performance - specifically larger context windows (200k tokens vs 120k for Claude 3.7) and expanded tool usage limits (200 vs 25 tool calls)
The default mode isn’t limited or nerfed - it’s the standard model implementation with normal context windows and tool limits that work great for most use cases. MAX mode is an optional upgrade for when you need those expanded capabilities, and are willing to pay the additional costs!
The Default option just routes to an available Premium model, but is no different than manually picking the model from the list!
4 Likes
Hey, but consider the official price list of Gemini Pro 2.5. However, the price of this model up to 200k tokens is much lower than Claude. This should be reflected in a fair use approach.
Input: $1.25, prompts <= 200k tokens
Output: $10.00, prompts <= 200k tokens
I think we should get at least a 200k context window, for this model.
In my opinion we have too big a gap between Premium and Max, there is nothing in the middle with a reasonable price tag.
2 Likes
Yes, cursor does not support the latest deepseek model and openrouter/quasar-alpha
Very sorry
2 Likes
In practical usage, users will ultimately be the judge - regardless of how things were in the past, how they are now, or how user expectations continue to rise in this era of technological explosion. Every day brings new experiences and higher standards. In fact, one reason Cursor received such widespread acclaim in its early days was precisely because it rode this wave of technological breakthroughs. Let’s wait and see what comes next. 
2 Likes
yes, but…
See my post history… BUT
The real prob with @cursor is the fact that it (THE TEAM) are completely disjointed from the community to the point that they do NOTHING to address issues such as issues with agentic necropsy (the agent is acting in a zombie fashion)
1 Like
MAX mode provides enhanced capabilities beyond default model performance - specifically larger context windows (200k tokens vs 120k for Claude 3.7) and expanded tool usage limits (200 vs 25 tool calls)
Cursor tells me that the none max sonnet is ‘less intelligent’.
Is the difference definitely only context windows and tool usage?
If yes, then I think the description should be ‘Less context and tool calls’.
If it is less intelligent in some other way, can you elaborate how?
6 Likes
If that’s the logic, why does Gemini have MAX Modelli? Is the Gemini MAX a 10 million context?
I agree, charging more is one thing, and not working is another
I think the most realistic strategy for devlelopers are:
- Use cursor for TAB autocomplete. This is really great which I haven’t found better alternatives.
- Use Claude 3.5 for relatively simple tasks which you MUST provide all the context required yourself using
@ - For any advanced models such as Claude 3.7 and Gemini 2.5, use
Roo Coderwhich is cheaper and much more controllable.
1 Like
No doubt this is the case. Since they have launched Max models, even Claude 3.7 with thinking does dumb things. Feels like the premium plan is of no use. I keep paying for Max usage with a lot of tool usage. If this is the case for long, Cursor will soon go down the drain
3 Likes