My university professor would always say that the real test is when you make your product available to the public… and Flash 3.6 failed (for now)…
Context
I have a project that I personally experiment with on a daily basis, with 99% of the development being done by AI agents. Why? Because I run it locally with no internet connection on VMs in Proxmox, so it’s basically very low-risk vibe coding to learn how and when to use AI models.
Google said that the 3.6 model has more quality, its faster and codes better. Let’s check the same benchmarks they use to validate this claim by looking at the metrics that show the differences between them:
1.FASTER? TRUE
2.CHEAPER? MAYBE*. I have to run more times when compared with 3.5
3.THE SAME? FALSE.
My first thought is:
“It seems that some tests from 3.6 are missing. Are they hand-picking metrics to make it appear that 3.6 is as good as 3.5?”
I don’t know if anyone has found a way to use this model reliably, but to me it feels like this model lacks what 3.5 has, and I would like Google to publish the missing benchmark data.
Google is losing when it comes to model comparisons, but it has a strong chance of winning if it creates a better GitHub and a better developer forum. The current one is generally not very good. If Google simply “copies” the best parts of Stack Overflow and GitHub, then combines them with its existing ecosystem, I think it has a strong chance of coming out ahead.
BTW, I don’t see why “gemini 3.5 flash cyber” was not released to antigravity users, bad decision when Google is losing in model quality when compared to Claude, GPT and Grok.
SOURCE:
After some similar experiences, I use 3.6 Flash only for some simple question & answer, not for Agent coding. 3.5 is much better in this regard imho.
Me too. I have been having issues with Gemini 3.6 because it does not seem to follow instructions as reliably as 3.5.
And the metric of this is missing!
WHAT A CONICIDENCE !
It often starts implementing features that I have explicitly decided to postpone because it is not yet time to “attack that front.” This can break the application, since it begins building functionality that the current environment cannot support due to hardware limitations.
For example, the application will eventually run on hardware designed to remain powered even during a grid outage, similar to a laptop with a battery. However, the system is still running in a VM, so no battery is currently available and the application cannot retrieve battery status or other related machine-state information.
Despite this constraint, the model starts developing a battery-monitoring module as though the required hardware already existed.
I gave up an migrated fully to Claude and Claude code. I may have doubled my output while having fewer problems.
Much smoother than antigravity with gemini
Have you tested Flash 3.6 and 3.5 via the API, in your own application, or using claude-code?
Because if you’re testing these models in Antigravity, it makes no sense; the tool is flawed and can completely ruin your tests.
API? do you mean google cloud services? or do you mean CLI ?
API, as in the pure model, you can get the keys from AIStudio. The point is that running tests in the Antigravity environment makes no sense, because that tool messes up outputs. Version 3.6 might just as well be better, but Antigravity degrades the model’s performance more severely. It doesn’t matter whether it’s the Antigravity APP, IDE, or CLI, they all suffer from the exact same issues.
I understand that the API is generally the better option for coding, and I already use it for more serious projects.
In this case, though, I’m referring specifically to Antigravity under the monthly plan. I mainly use it for personal projects, lighter coding tasks, smaller projects, or individual small modules within a larger codebase.
The main point is that we’re discussing the Antigravity ecosystem, not the API, since they are separate platforms with different billing models.
My main feedback is that, within Antigravity, I currently find 3.5 better than 3.6. This post is meant to share my experience and thoughts about the new model specifically within that ecosystem. If 3.5 is removed and fully replaced by 3.6, I would consider that a downgrade to the tool in its current state.
