When it comes to 3.7, how I felt it in google antigravity is it’s not accurate and make mistakes more than 3.6 and 3.8. So it’s not reliable.
Gemini 3.8
It’s actually a good model, but in my experience in antigravity IDE is that it reads all the related files again and again fully wasting all the quota. It’s accurate and performance is good, but in my opinion it wastes the quota and overdoing things. Even with UI designing, it does more than needed. So quota wasting is a big problem.
Gemini 3.6
This is the balanced model I would say. The good thing is that it makes changes quickly without reading a lot of files or reading them fully. It’s like it quickly knows which file’s which code block to updtae. It’s also fast. Accuracy could be a little isssue with it sometimes but it is random. But overall I prefer the Gemini 3.6.
I hope my opinions would help someone feel good and identify the models more. 3.8 is a good model for identifying bugs caz it reads the whole file. 3.7 is IDK it’s really not reliable. But 3.6 is really like the best yet.
I’m sharing this so everyone can understand how different models perform and to not make anyone unhappy. It’s just my opinion, so hope you all would be able to develop things much easier.
I have been using 3.8 since it became available and I do agree that 3.6 feels subjectively better and is objectively faster. I use 3.8 High for planning and 3.6 Low for the implementation.
I have to disagree on 3.6 being the better model. 3.8 has been completely amazing for me.
It is the first Flash model I actually enjoy using. Before 3.8, I was constantly having to fall back to Pro because the older Flash models just could not handle complex tasks. Now, 3.8 gets them perfectly and I almost never need Pro anymore.
It sounds like the IDE integrations have some annoying file-reading bugs to work out, but the model’s actual reasoning is a massive leap forward. Seeing people ask to go back to 3.6 worries me a bit. I really hope Google does not roll things back just because of some Antigravity quirks.
That’s exactly what I’m doing too. 3.8 is better for planning and when you need a high quality UI or bug fix. But subjectively 3.6 balances everything. We can fix a bug with 2 prompts at a time if something goes wrong with 3.6 but with 3.8 it eats the whole quota with just 1 prompt which we could do with 3.6 with maximumly 10% of the quota.
No actually you are corerct! 3.8 is great, but for developers who have paid plans. Caz 3.8 uses a lot of readings and tokens and produces an amazing output as well.
When it comes to developers who have free plans actually 3.6 does work quickly and with less file readings and time. The quality cannot be compared of 3.8 and 3.6, but the overall metrics consider how they work for majority right.
We should not go back to 3.6. But 3.8 definitely needs more improvements. The thing is my post is not asking to setback 3.8, it’s more of a comparison of overall metrics.
