This is happening because the CurL command calls Gemini 1.0 pro which is a deprecated model. However, even if the verification fails, it still succeeds to add the api key. This is the result of my testing. However, I’m not able to add Gemini 2.5 pro exp model either. Others work.
It’s a Ui bug mostly on cursor is what I believe.
Does it work in agent mode ? Any LLM with function calling support available should be opened for agent mode. There seems to be a bias from cursor team to allow only specific models even though others support function calling. Ex. Gemini 2.0 flash also supports function calling but isn’t available on agent mode. Please @cursor team, we are paid users and you can’t restrict on what you had to on even paid users with pro subscription.
Does this support agent mode ? If not, there’s no point. I’d be happy to copy paste it from aistudio itself rather than using the ask mode. Why do you I need a pro subscription to just copy paste and not leverage the power of agentic coding ?
Cursor has added it already.
Also to add a model, paste model name and press enter. Dont click add model.
Agent mode should happen as this Gemini model is really good.
But you can use edit mode and it’s decent than using AI studio.
I still feel Claude 3.7 sonnet thinking is better, maybe because it’s already adapted for agent mode, but it’s getting there
Actually Gemini Pro 2.5 already works well as an agent. One just has to ignore the warnings about it being unsupported as an agent. Then it works fine.
Yes it works but not as good as sonnet yet. Maybe cursor needs to optimize it in agent mode.
I have Cursor 0.48.2 and Gemini 2.5 is one of the available models now. It shows up in all modes.
You of course have to turn it on in the settings first though.
gemin 2.5 pro doesn’t seem to show its thoughts on Cursor? Is that expected?
@deanrie
Thank you for adding Gemini so quickly. Two questions:
I appreciate your attention.
I just witnessed a kind of hallucination by Gemini-2.5-pro-exp-03-25, like no other.
It simulated everything => code change, code execution, test runs, and command executions in shell, and all the tests passed with 100% coverage, without any of it being true.
I am in awe… never seen anything like it
and then it became non responsive with this message:
" We’ve hit a rate limit with gemini-openai. Please switch to the ‘auto-select’ model, another model, or try again in a few moments.
(Request ID: 4f38c574-0d81-46be-92fa-367942023f6a)"
amazing simulation
Interesting, I have experienced such hallucinations with all most capable models from all providers. Sometimes AI pretends to search online and fakes a tool call.
Its obviously not calling anything as it just outputs the headers
quite fascinating.
Please, we need an answer. Will the developers optimize cursor for agent support? Currently, cursor performance is very poor for agentic use with Gemini. Please, this is a request: don’t focus your product on just one AI (i.e., Claude Sonnet). If in the future they increase their API price from $15 to something similar to what OpenAI is doing with their recent version, you will lose all your customers and face significant financial losses as well.
its better than sonnet?
Yes, for some tasks like analysis and planning, actual code is not always better.
So it looks like agent mode (tool calling) works when using the Cursor proxy, but when using my own Google API key, I just see “tool_code” blocks that are not executed.
EDIT: nope, totally inconsistent. I just had one tool call work, and then another fail, right in the same conversation, using the Cursor model proxy.
It works for me. But doesn’t support “vision”. But the actual model does…
This model is so good that I can’t wait anymore
