You legend. That was it. Thanks. Just hit return don’t press add model.
I compared Sonnet 3.7 vs Gemini 2.5 exp pro in Cursor.
I have a task, in short: “Create the same searchable dropdown on X page like we have on Y page. Use the same dropdown values that we have on Y page.”
Sonnet 3.7 - very well (Agent mode)
Gemini 2.5 - first attempt - created a new page instead of using existing - not very well (Edit mode)
Gemini 2.5 - 2nd attempt - the same prompt, but it did not get the task at all - poor (Edit mode)
My conclusion:
We should definitely wait for Cursor optimization changes to force this model work effectively
5 Likes
omg, I feel dumb now - thank you sir!
The Agent mode provides basic planning and practice, and will automatically correct errors, which is very important for coding. The Agent mode is crucial, whether using DeepSeek or Gemini.
Were you able to use Agent mode with Gemini 2.5? When i send a prompt via Agent mode it can stuck or never generate any result.
I’ve been trying out Gemini 2.5 Pro and it’s really good. Just miss the agent mode. Is there any way to use it with an agent like setup? I’ve been really enjoying claude-code - and I’d wish for a claude-code with gemini 2.5 pro!
The Cursor team would have to check the model and test it to ensure it works well with all the agent features.
Plus also when many users want to rely on it it should not be like when 3.7 was launched. Overloaded and breaking for everyone as whole world was rushing to use it.
I guess a bit patience until its usable for Agent
1 Like
Well they never ended up supporting Agent mode for previous gemini models (which were pretty good).
I think the model best suited to agents still is sonnet 3.7 - but something like Aider’s architect mode would be great - use Gemini for large scale thinking AND for editing files/generating code. But use sonnet 3.7 for tool calling
1 Like
Yea exactly
I added Gemini 2.5 Pro without any issues. Just enter your own Google AI key and add “gemini-2.5-pro-exp-03-25” model to the list manually. Ignore “key verification failed” error - Google API will work anyway.
3 Likes
Yes able to add and test it eventually. Now we need to wait for Cursor to support it in Agent mode. Maybe this time they will add.
they have to i tested this model and coding is by far state of the art also it will be 100% cheaper the sonnet so adding this model is no brainer
Been trying 2.5 via OpenRouter for just couple of hours.
I’m a bit skeptical that G2.5 will work fine in Agent mode as it doesn’t seem to grasp function/tool calls even though it should see what Cursor offers. Maybe with one of the next updates (or new model), Google will offer a real agentic model.
Besides the above: G2.5 really seems to reason way better than Sonnet 3.7, gives plausible reasoning (though lenghty) and does more concise code edits. But if it deems some method to be “useless”, it’ll remove it, too.
Thanks man.
Just tried this and seems to work, thank you (Latest Version + Ignored the error)
Took some time to generate the first response but it’s doing it!
Just made a video on how to integrate Gemini 2.5 Pro App Build With Cursor AI - Is It the Best?! Thanks to @iamserjo and everyone else for the input here
Gemini does not beat Claude 3.7 on coding tasks. BUT it is SOTA on large context understanding, needle in haystack, and on math and reasoning
So Cursor devs. Please look deep into my proposal:
Split planning and coding into 2 things. This is requisite for different model (defaults or user choice) used for them.
- Gemini 2.5 pro can be the planner. Give it whole codebase as context if not more than say, 500k tokens (half its current max). RAG is now holding back Cursor Agent so unshackle it, when Gemini is showing its best at large context and economically so, by a mile compared to other models
- Claude 3.7 can remain as implementer, according to plan and code files, lines identified by planner the coder should implement changes in
How to ping devs on this? @deanrie can you help
5 Likes
I always got Unauthorized API key, even I use correct one from aigoogle
@Cursor Team, in addition to agent mode, is it possible to make work Gemini as a MCP client? (With workaround)
you dont need to use API key. cursor has added gemini 2.5 pro now.