Diverted to Cursor-Small On Slow Requests

Discussion
1

After using all the fast requests, now instead of slow premium requests I am often getting diverted to cursor-small (in Agent mode) over the last few days when trying to access Sonnet. Obviously I’m not the only one, as many have noted stalling and dumb responses that ignore the project structure, docs, and duplicate files etc.

image
image833×316 23.1 KB

If it’s the Anthropic API hitting rate limits or something, we need some feedback as to what is wrong and how to deal with it. If it’s abuse/bots overloading the API or keys getting revoked by Anthropic or some other unforeseen cause it would be nice to know. In which case (feature request:) there should maybe also be some kind of alert that lets users know there is an API issue with Anthropic.

I also understand API costs are a problem, but if this is deliberate on the part of Cursor, it needs to be made more clear that slow requests are not guaranteed and may be diverted to cursor-small.

P.S. Cursor is such a great product (Thanks devs!) with Sonnet that without Sonnet working, it’s almost like we have discontinuation syndrome as many of us have found it quite indispensable! That’s how good a product it is.

Thanks.

2

Hey, as Sonnet is selected as your model at the bottom, it will be Sonnet you are talking to, even if you are using slow requests! We never downgrade the model you are talking to without telling you, and that only happens if one model is unavailable (due to downtime, etc).

Additionally, cursor-small cannot be used in agent mode, so it would have to be a different model you were talking to, even if the above wasn’t the case!

What makes you think you aren’t talking to Claude here?

3

Is that supposed to be a joke? Can you not see the response or have you never used Cursor/Sonnet? Claude Sonnet does not call itself cursor-small or think it was designed by Cursor. It stated it is designed by Cursor during slow agent requests when Sonnet is selected, in agent mode. What makes you think it’s Sonnet when its performance clearly confirms it is not?

When I posted an image of it stating it was a Cursor model and showing that it reported itself as a Cursor model, my account was flagged and the post was hidden. Now you say that’s Sonnet? Seriously? My small model requests counter is going up when Sonnet is selected in agent mode ffs.

Your insistence it is using Claude Sonnet in the face of numerous reports of Sonnet behaving like it got hit over the head, suggests you are either in denial of what is being reported (and thus perpetuating the bug rather than addressing it), or are are fully aware and are gaslighting customers.

Claude Sonnet understands the codebase and does not say that it is Cursor Small or think it was designed by Cursor.

It does not duplicate files and do all the dumb things people are reporting.

Sonnet slow requests in agent mode do/should not result in small model request usage counters increasing.

A slow request to sonnet in the normal mode is slow.

Slow request to Sonnet in agent mode are fast and about 50% less intelligent.

CURSOR IS NOT CALLING SONNET FOR SLOW AGENT REQUESTS FOR ME. THIS IS NOT A DEBATE. THIS IS NOT THE API BEING DOWN. IT WORKS IN NORMAL MODE. THIS HAS BEEN CONFIRMED, REPORTED WITH SCREENSHOT SHOWING IT’S REPORTING ITSELF AS CURSOR DESIGNED= CURSOR SMALL, AND THE POSTS REPORTING THIS WERE HIDDEN.

The AI knows who designed it and in agent mode if Claude were being called, it would say this instead of the above:

image
image832×107 14.5 KB

5

Hi, we tell whichever model you use, including Claude 3.5, that it is an AI coding assistant inside Cursor, within the prompt we send to it.
Hypothetically, all the models in Cursor should (although we don’t guarantee) reply saying they are AI coding assistants designed by us.

I tell you these statements as fact:

  • Excluding rare occasions when a LLM provider like Anthropic is facing downtime, you will always get the model you choose responding to you
  • cursor-small cannot be used in Composer’s agent mode - it does not support it, even if we we’re redirecting you to another model.

Even with premium models like Claude 3.5, they are black boxes at the end of the day, and their behaviour in certain situations can be drastically lower than others. Coding simple Python is easy for Claude, but coding obscure or new languages and technologies is much more difficult, and would return much worse results.

While we try to supplement that with the @web and @docs functionalities, as well as the context of your codebase, there is still model-level limitations we cannot yet overcome entirely, which is often the cause for when Claude doesn’t seem to perform like itself sometimes.

I hope this explains things out, but if you have any further questions, let me know!

6

Thanks, but that explanation doesn’t quite explain the often (but intermittent) behavior we’re seeing.

1: I am talking about things that it could do before, suddenly it is often not capable and the only difference is we are out of fast requests; not increased context, but failure to understand the same context. This is consistent with other recent reports.

#2. Obscure code and languages are not going to cause Sonnet to go from capable of during a fast request, to GPT-2 level. Whatever the cause, it’s not me, my code, the Anthropic API, or the model. All of that is confirmed to work fine outside this failure mode.

#3 The context is not the problem as Sonnet can understand the (same) context during the fast requests, or when it’s not in failure mode(peak hours?). It absorbs the context masterfully, outside this intermittent failure mode.

When it is capable, it consistently reports that it is Sonnet. It does not during failure mode. I have never had it or GPT4o etc. identify as a Cursor product either for what it’s worth. The point being that if it reports as Sonnet I go on to use the product without issue. If it reports as a Cursor product I have learned it is not capable, and it’s performance is consistent with cursor-small even though it “can’t” be used in agent mode and is not selected. GPT4o, the other agent mode model, doesn’t exhibit any of this behavior either.

I’m not the only one noticing that it worked with fast requests, and now suddenly the next day is trying to make changes to files that don’t exist and any number of other failure modes.

I believe I saw at least one report of this going away when using one’s own API key.

There is no downtime in the API as Sonnet works fine in normal mode as far as I can tell even during this failure mode in agent mode.

*Looks like a duck/bug
*Quacks like a duck/bug
*Walks like a duck/bug

Hunt. The. Bug. Or refund all customers and close shop or sell to someone who hunts bugs.