Yes, it’s clearly much better. But it’s not faster, and the usage limits seem a bit more restrictive.
Overall, the usage limits are acceptable and quality is much higher, so it works for me
It all seems to be going backward with each new model they produce. The new version of AntiGravity is now unusable because the current ai versions are so poor. Claude works well but it’s so limited - without the ultra tier - that it’s not usable in any expanded capacity.
I suspect that the model training data is getting poorer. They’re using companies like Outlier to help ‘train’ the new models and, having been a trainer for 9 months on that platform, I can truly say the level of output is disgustingly poor. People aren’t given enough time to complete training tasks thoroughly enough. Most are rushing to complete the tasks in order to get paid - because payment is dependent on task completion within an overly tight timeframe for many projects. It’s a massive problem. Also there are many people on those training platforms who are trying to earn a quick buck and willing to cut corners doing so and still others who have no clue about ai or what they are supposed to be doing. The vetting is a serious issue.
Only Gemini 3.1 Pro can still solve my problems — it’s the only model whose responses I can trust. Gemini 3.5 Flash ignores instructions and its answers are full of hallucinations, making it unreliable. Yet Google has significantly restricted Pro’s usage quota ever since the release of Gemini 3.5 Flash.
It’s just gotten faster, more expensive, and buggier. Honestly, it’s a worse experience and a real step backward.
If this is the AI that is going to take over the future and put humans out of work, we can all relax. lol this thing can’t reason its way out of a cardboard box. 99% failure at tasks it excelled at in 3.1. Shocked they would release such nonsense when they were on a roll.
Hey All,
Thanks for all of the feedback on the model. We’ve rolled out a new version of Gemini 3.5 Flash in Antigravity that boasts much less and has higher endurance on harder tasks.
Is there a way for models to come with a “release version hint”/changelog? if not possible, what about a blog post? having to find “changelog entries” in random posts isn’t great. Thanks.
In my estimation, the fundamental grievance resides not within the cognitive constraints of the model itself, but in the opacity of IDE prompting coupled with a cascade of systemic vulnerabilities. Consequently, the AI charges forward heedlessly without verification; rather than disclosing discrepancies, it endeavors to resolve them in isolation. Even amid operational failures, it spirals into an infinite loop, effortlessly depleting one’s entire quota while exacerbating the underlying issues. Simply put, they conceive of AI agents not as tools or workflows, but as outsourced subordinates—save for the fact, of course, that there is no recourse when they breach their covenant. How alluring this must sound to decision-makers divorced from active development; indeed, actual developers are not their intended audience at all.