Gemini-2.5-pro-preview-05-06

Does Gemini support MCP yet? Frustrating that it hasn’t.

Man, what a fierce fight! Our ChatGPT is scrambling to keep up! :face_with_head_bandage:

This model has the tendency to say the words “Please try running the script again.” I guess these were highly upweighted during post-training. This stops it from acting like an agent that autonomously updates a script, tries to run it, tweaks it, tries to run it again, etc. I’ve tried to tell it to run an autonomous debug loop but it really wants to say exactly “Please try running the script again.” rather than follow my instructions. I can work around it by asking several times in the same prompt. Perhaps something the cursor team could add to the system prompt?

Yep


I am getting sick of it just doing 90% of the work and then giving me a “task” to do that its perfectly capable of doing

@the1dv do you include in your prompt or in your rules the instruction to make those adjustments?

some of those new and updated models are very literal, so if you do not ask it to make the change it wont

Following will do the change
Make the adjustment....
OR
Adjust code to....

Following will not make changes but tell you how you can do it
How can this feature be changed to...
How can we make this work?
Can you do this differently?

Gemini calls the tools within Cursor just fine

1 Like

image
image

Yes - I am literally asking it to make the changes and when it actually connects it tries and then it fails in some very interesting ways - constantly

This is the latest one “Oh it matches what I want”
It literally doesnt.

To be clear - this is not a model problem - this is a tool problem with how cursor have implemented the file_edit tools and diffs all in the name of saving money
Using dumber models that allow up to 3 failures to apply (aka 3x the cost to us) before reverting to the model it was previously working is disingenuous.

image

Ive restarted - it just doesnt work with Cursor for me unfortunately - its a shame because the actual suggestions seem bang on and if it applied it would be a fanstastic model for this.

That is very strange indeed. Please create a bug report and add the info you provided here as it might need to be looked by Cursor team. Add to the bug report this info. Also there include a Request ID from the chat that fails. If Privacy is set to on, please try one edit (even if on different project) where you can reproduce the issue and then attach the Request ID so the team can take a look at what happened, you can turn on Privacy right after again.