Gemini 3 Pro often just stops in the middle of a task or runs into errors endlessly. I really don’t understand how, after all this time and so many upgrades, they still can’t fix such a basic problem.
It seriously makes me question whether Cursor is deliberately wasting users’ tokens this way.
same here
in my experience Flash performs better than Pro. I also tested in Antigravity and Gemini 3 Pro still sucks there
its not very good, good maybe at more tasks other than coding.
Just curious, have you tried other models than Gemini? My own experience is that Gemini usually comes in last, compared to other models. Most of the others are much more reliable. I’ve used Gemini, as well as GPT-5.x, Grok Code, Composer, and the ones I use most are Sonnet and Opus these days. The latter two are of course very expensive, but the team I am on generally uses those exclusively, or just use Claude Code, and we have had great luck with them overall.
If the expense for the Claude models is too much, I’ve had pretty good luck with Composer and Grok Code Fast. Its been hit or miss with the GTP models for me…some people seem to have great success with them. For me, the ridiculously long thinking cycles (which often take a minute or so) suck up time and put me off, but I have also run into cases where GPT-5 models will just stop and bail, even though I know they could continue and finish the work. GPT-5 models, though, have still provided consistently better results and quality, than any of the Gemini models in my experience.
It may just be worth trying out a different model, to see if one of them better suits you, how you work, the kind of work you do, etc. Not every model is well suited to every kind of software development work, too. For that matter, different models seem to do a better job than others at specific kinds of tasks within the software development arena…planning, design & architecture, implementation, specific languages and/or platforms, specific third party libraries, etc. Sometimes its just better to switch models and see if one handles your tasks better than the one you might initially prefer.
Gemini is know for Hallucinate more. its on Models that base on Benchmark. Thats why base on my experience, i never use Gemini Pro 3, and use only Flash 3 for execution. Plans?
i use GPT 5.2 or Sonnet 4.5
谢谢你们的分享，但是根据我的经验，如果你尝试完成一个小范围的编码任务，其他模型或许能做到更好，但是，如果你的任务涉及比较多的上下文信息，我觉得 Gemini 3 pro 会有更好的理解能力，而理解能力在我看来是更重要的。
我曾经尝试用 Claude sonet 4.5/gpt5.2/Gemini 3 flash/Gemini 3 Pro来解决同一个比较复杂的重构任务，最后只有Gemini 3 Pro准确理解了我的意图并且以“合理”的方式完成了任务，AI在编程中（当然也包括其他任务）最难的是做到“合理”，“合理”的意思是——感觉就该这么做，这相当不容易，有赖于模型对复杂上下文的理解能力。
有人说 Claude sonet 4.5 善于编程，但我一直有一个观点：编程不仅仅是写代码，更重要的是“理解”，比如：
这就是我更信任 Gemini 3 的原因。
有兄弟提出 Gemini 3 flash 比 Pro 更好，我同意，但不是所有任务都更好，Gemini 3 flash有时候会偷懒，就我的体感来说，完成复杂任务，Pro 一次性成功的概率会比 flash 更高。
Gemini 3 flash 我觉得比 Pro 更好的点在于：flash 不会过渡思考，它像 pro 一样聪明，但更“心直口快”，这在有些任务上会更符合你的直觉。而 pro 则更“深思熟虑”，它提出的看法可能并不符合你的第一直觉，但这并不是缺点，反而，有时候不符合我们直觉的原因很可能是我们“太菜了”，一时半会没有看明白高手的深意。
Thank you for your sharing. However, based on my experience, if you are attempting a small-scale coding task, other models might perform better. But if your task involves a significant amount of contextual information, I believe Gemini 3 Pro has a stronger comprehension ability—which, in my view, is more important.
I once tried using Claude Sonnet 4.5, GPT-5.2, Gemini 3 Flash, and Gemini 3 Pro to tackle the same complex refactoring task. Ultimately, only Gemini 3 Pro accurately understood my intention and completed the task in a “reasonable” way. The hardest part of AI in programming (and other tasks, of course) is achieving “reasonableness”—it just feels like the right way to do it. This is quite challenging and relies heavily on the model’s ability to comprehend complex contexts.
Some say Claude Sonnet 4.5 excels at programming, but I’ve always held the view that programming isn’t just about writing code—it’s more about “understanding.” For example:
Understanding the goals of your project
Understanding your intentions
Understanding the real-world implications of the code (such as better grasping the spatial structure of a UI)
These abilities actually extend beyond mere coding. That’s why a model with more parameters and multimodal capabilities typically performs better in this regard than a purely programming-focused model.
This is why I trust Gemini 3 more.
Some friends mentioned that Gemini 3 Flash is better than Pro. I agree, but not for all tasks. Gemini 3 Flash sometimes cuts corners, and from my experience, Pro has a higher success rate in completing complex tasks on the first try.
Where I think Gemini 3 Flash outperforms Pro is that Flash doesn’t overthink—it’s as smart as Pro but more “straightforward,” which can align better with your intuition in certain tasks. Pro, on the other hand, is more “thoughtful.” Its suggestions may not align with your initial intuition, but that’s not a flaw. In fact, sometimes the reason it doesn’t match our intuition might simply be that we’re “not experienced enough” to immediately grasp the deeper insight behind an expert’s approach.
Hey there, do you know if cursor team is actively working on making the harness workable with Gemini 3 Pro?
Claude models works well everywhere I’m assuming Cursor team started out with claude models - I’ve seen them use Claude in most of their video demos. GPT sometimes only.
I’m just wondering if a model works bad - maybe some configuration could make it atleast 10% better - Google has provided this document for prompt engineering gemini 3: https://ai.google.dev/gemini-api/docs/prompting-strategies and there’s a lot in there that cursor could use to modify their model harness for gemini 3 pro that could make it better imo.
Thoughts?
