Sudden and Severe Changes to Model Usage Limits – Impact on Paid Ultra Plan Users

I’ve been using Google’s AntiGravity editor for software development, actively working with Gemini 3.1 Pro and Claude Opus 4.6 models under the Ultra (second-tier) subscription plan. I am currently in my second month.

During the first month—especially at the beginning—the experience was smooth. Usage limits would reset quickly, and it was practically impossible to hit the limits during normal development workflows. This made the platform reliable and productive.

However, toward the end of the first month, there was a significant and abrupt change in how usage limits were handled. Since then:

  • Usage limits have become extremely restrictive.

  • The previously implied “refresh every ~5 hours” mechanism no longer works as expected.

  • Instead of resetting automatically, the timer seems to start only when I begin using the system, which creates inconsistency and unpredictability.

  • At this point, I am often unable to use Claude models at all.

For example, as of now, the system indicates that all my limits were exhausted yesterday and that new limits will only be available the next day. Previously, even if limits were tight, I could at least send a few prompts—but now, even that is no longer possible.

The core issue is this:
I subscribed and continue paying for this plan based on the capabilities and limits presented at the beginning. Those conditions have materially changed, and the current service no longer reflects what was initially experienced or expected.

While I understand that terms of service may allow changes to limits, unilateral and substantial reductions in usability—especially after users have committed to a paid plan—create a serious reliability and trust issue.

Additionally:

  • I’ve noticed many similar complaints from other users, particularly around Pro plans, where limits are reportedly exhausted after just a few prompts.

  • I am using one of the highest-tier plans available, yet facing the same issues.

  • I have not been able to find any effective support channel. Feedback submissions have not received responses.

Expectation:

  • Clear, transparent communication about how limits actually work.

  • Consistency between advertised/initial experience and ongoing service.

  • A review of current limit policies, especially for paid tiers.

  • A proper support channel with real responses.

If the infrastructure cannot sustain the promised service levels, then these plans should not be marketed in their current form.

At this stage, this is no longer a minor inconvenience—it is directly impacting productivity and creating a situation where the service is not usable as intended.

I would appreciate this being taken seriously and addressed with a concrete resolution.

It’s gotten very bad in the past week or so. For me, it takes about 6 messages to completely drain my Claude use. A couple weeks ago I could use claude for about 4-6 hours per day.

On top of this, I randomly get my workflow cut in the middle, and a message about server issues (that tells me to try again in one minute), but it still uses my tokens. Additionally, when I get the server error, it seems to use more tokens than a normal message would of.

Unacceptable and unusable.

Also, Gemini (every model) has gotten far worse than a couple weeks ago, and far worse than a couple weeks before that. It is making huge mistakes on a regular basis. I literally have to have Claude spend its entire token limit to audit and fix Gemini’s mess.

RIght now the only option I have is to not use gemini at all, and use claude fresh, then it can last a tiny bit longer, but with the current token usage that’s no where near worth the subscription.

Whatever is going on is a total joke. Not to mention the language server wrapper has been getting worse with the last two antigravity updates, hogging insane resources, forcibly scanning directories it shouldnt be, and forcing the entire directory into my RAM and skyroketting my CPU usage.

I have telemetry off. The cloud AI scans my local files and chats for data (can barely acces anything on the server side relating to my chat logs), why is language_server pushing my entire work directory to the cloud?