Paid Ultra Account Auto-Banned

Google Antigravity
1

Hi everyone, and specifically calling on Google/Antigravity Community Managers, Employees, or Product Experts (like Fred SR):

I am posting here out of sheer desperation. I am a solo developer who pays full price out-of-pocket for the Ultra tier. Today, my account was abruptly suspended for alleged “policy violations,” and I am 100% certain this is a massive False Positive triggered by the platform’s own backend bugs.

Here is the exact sequence of your system failures that led to my unjust ban:

1. The “Free-Tier” Downgrade Bug: As confirmed by Product Experts in other threads, Ultra users are wrongly being subjected to free-tier limits. I was hit hard by this, constantly hitting an undocumented ~1.5M tokens/hr hard throttle on claude-opus-4.5-thinking, resulting in a flood of 429/503 Capacity Exhausted errors.

2. The 20k Context Bug & MCP Crash Loop (The Trigger): Worse, my Ultra account was severely misprovisioned with a hard 20,000 token context limit instead of the standard 200k. Because my coding workflow relies heavily on MCP (Model Context Protocol) tools, the mathematical reality is that [MCP Tool Schemas] + [Claude's Internal Reasoning Traces] instantly exceeded this tiny 20k limit.

Because the model physically lacked the context buffer to execute the tools, my legitimate API calls were constantly crashing, returning truncated errors, and forcing my client to automatically retry.

3. The Automated Bot Misjudgment: It is painfully obvious what happened next. Your automated Trust & Safety bot saw this rapid succession of bug-induced 503/429 errors, failed tool executions, and client-side retries. It mechanically and blindly flagged my account for “malicious botting,” “API abuse,” or “abnormal traffic.”

I am essentially being punished and banned by your security bots for the architectural failures of the Antigravity platform itself.

I have strictly adhered to the TOS. No account sharing, no scraping—just a solo dev trying to use the service I paid for, which was completely broken on my end.

My Request: Email support is currently a black box of automated rejection templates. I urgently need an actual human employee to manually review my logs. A real engineer will immediately see that the “spam” was just my MCP tools choking on your 20k context bug.

Can a Community Manager please help escalate this to the Trust & Safety team? I am completely locked out of my daily work.

3 Likes
2

As a strong supporter of the Antigravity vision, seeing a case like this is genuinely alarming for the developer community.

Referring to the docs/plans, parts of this service operate under Preview/Pre-GA terms. In a Preview environment, system instability—like the context crashes and 503 loops described by the OP—is an expected engineering reality.

The critical failure here is the mismatch between the product stage and the enforcement logic. It is logically unsound to apply production-grade, zero-tolerance automated bans to a Preview-grade environment where the platform itself is likely triggering the “abnormal traffic” via internal bugs.

By penalizing paying users for platform-side architectural failures, you risk dismantling the trust of the early adopters you rely on. We urgently need a “human-in-the-loop” review process for these technical edge cases, rather than black-box automated rejections.

3 Likes
3

I’m going to tell you that you’ve already admitted to using disallowed third-party software by being able to deduce which HTTP errors were provided by the backend. This is utterly impossible in a stock, unmodified Antigravity install. You might want to consider using a stock, unmodified Antigravity with no extensions that enhance the Antigravity experience in any way, shape, or form, as annoying as that may be.

I’m going to take a gander and say you were using Antigravity Claude Proxy with Claude Code. Sucks that this is disallowed, but it should probably be the only “unofficial” client stack allowed as it’s in the spirit.

Unfortunately, I have reasons to believe that strict controls on third-party software are done to ensure uniformity in the model completion requests (same Antigravity system prompt, etc.) in order to create additional training data for future Gemini models. The Antigravity Terms of Service explicitly states that your sessions may be used to train AI models.

There’s more I’m thinking on this, and all I’ll say further is that GPT-OSS 120B is only offered as a MacGuffin.

1 Like
4

First, Antigravity is an experimental product, largely due to the departure of a core R&D team from Windsurf, as we all know;

Second, starting from the HTTPS request, we understand that while Antigravity sends model requests to Vertex, it injects code within the process to weaken the model’s inherent capabilities.

From the day Antigravity launched until now, I’ve been using it extensively to develop my projects. I can tell you with absolute certainty: Antigravity is currently completely incapable of supporting the development of any project specification.

Here are just a few critical reasons:

  1. It truncates model outputs and task specifications, making it impossible for any task to run in a loop for more than 20 minutes—utterly unsuited for complex document design, project exploration, module development, microtask design, testing, and countless other scenarios.

  2. The output context is extremely narrow, making it impossible to produce satisfactory planning results or code modification outcomes!

  3. Calls to MCP are extremely lazy and inadequate—this may indicate a severe issue with command adherence.

In fact, from the release of Claude 4.5 Opus until now with 4.6 Opus, I’ve only experienced the true power of these models when using official APIs within Claude Code, Cursor, and OpenCode.

Similarly, only by integrating Gemini series models into IDEs and CLIs like OpenCode can I fully leverage the models’ inherent coding capabilities.

HOW PATHETIC THE GOOGLE USER WE ARE!

So the question arises: If Antigravity prevents me from working properly, why should I, as a paying user, be barred from finding my own solutions to access the corresponding services?

Neither Cursor, Windsurf, Claude Code, Codex CLI, nor iFlow impose such stringent service terms—only Google does!

5

Look, I agree with you, but it is what it is, at least in the moment. I wholly disagree with you about this position.

Looking deeply at the Terms of Service, I think the relevant line they might be hitting us with is this from Google Cloud Service Specific Terms, 16(c):

c. No Reverse Engineering. Customer will not, and will not allow End Users to, reverse engineer or extract any components of an AI/ML Service, Software, or its models (such as using prompts to discover training data). Google may immediately suspend or terminate Customer’s use of any AI/ML Service based on any suspected violation of the preceding sentence.

You can find the above subsection buried in the mountain of legalese here: Service Specific Terms  |  Google Cloud

Sufficient heavy use of third-party coding software may resemble a distillation attack, and Google has in fact mentioned recently being the victim of a potential distillation attack.

The Antigravity ToS has this section as well:

You must not abuse, harm, interfere with, or disrupt the Service. This includes, but is not limited to, using the Service in connection with products not provided by us.

Which doesn’t explicitly state that third-party software is prohibited. High and heavy usage that is otherwise normal and within the spirit and scope of the service provided is likely to count on its own as “abuse” because it was done with a third-party tool. If I read this correctly, it’s not that you used a third-party tool alone, but that you used that third-party tool quite heavily (which was likely flagged as “abuse” with the potential for “disruption” by consuming limited system resources.)

My interpretation of that clause was: “Third-party software is fine only if it behaves like an approved client.” If you start hammering their API with your third-party software, that’s when they’ll act.

Truth be told, I don’t actually know and I’m just trying to make heads or tails of the multiple Terms of Service documents they’ve provided us for Antigravity (the Antigravity ToS itself, plus all of the additional ToSes that the Antigravity ToS requires you to also accept.)

1 Like
6

I really don’t understand why you’re defending Google so fiercely.

Did Google pay you? Or did it make extraordinary efforts for your AI services?

Since you’re so good at crunching numbers, why not calculate the relationship between USERS and Google?

USERS give Google money, provide training data, and painstakingly tweak their projects to fit Antigravity or even Gemini Code Assist.

Fine. So what does Google give USERS in return?

Beyond delivering an utterly terrible model experience, Google employs brutally aggressive and unfiltered tactics to block users from improving their own experience in any way possible.

What possible justification could you possibly have for defending such a monopolistic behemoth—a trillion-dollar corporation that capriciously toys with user experience?

And you even resort to the laughable tactic of citing Google’s terms of service. So what? What good does that do? It’s like me posting a sign in my own home: “Anyone entering this house must accept a beating from me without complaint or compensation.” Does that make it legally binding?

Any rational person, especially those who genuinely use AI, knows that major players like Codex, Claude Code, Cursor, and Windsurf tacitly allow users to freely utilize their quotas to enhance their own experience. Only Google has been training its users.

The person I’m conversing with right now is a thoroughly trained slave of Google.

You’re truly pitiful.

1 Like
7

Why are you bothering with Google, then?

1 Like
8

If you think I am in any sense of the word “defending” Google, you are horribly mistaken.

I laid out the particular clauses of their Terms of Service that they’ll likely cite in this situation. I also noted the vagueness of one of them and how it’s open to interpretation. While I didn’t say it so directly, bluntly, and plainly as I am now, I even told you that it’s likely acceptable within the Terms of Service to use third-party clients if you didn’t use them heavily and kept it to agentic AI tools (fitting the spirit of the service.) The language doesn’t explicitly say that’s prohibited, only if the service was “abused” or such with those tools.

There’s also the section I mentioned about distillation attempts because it also relates to heavy usage of the API. The company recently had a high-profile announcement about a potential distillation attack in the news just four days ago - Google: Gemini hit with 100,000+ prompts in cloning attempt

That’s not defending Google one iota. This is explaining what logic they’re likely to use when addressing your situation, so that you can prepare yourself.

I personally do not believe that third-party tools are themselves banned in the Terms of Service and I also believe that your account disablement for using third-party tools may have been improper.

1 Like
9

Welp.. this will be my last year of paying for any service from Google.
$500 in the last 45 days from me.. Not next month tho :slight_smile:

10

I don’t have a problem with there being rules, (though if they conflict with my planned usage, then obviously I may shop elsewhere), but no one is going to read the TOS.

Prior to banning the service they should have issued warnings. Rate limit the account for 3 days and send an email to the user with an explanation, and guidance on how to correct the violation.

Your account has been detected using third party tools which violate our TOS. Your AI access has been suspended until X. Please correct these issues to prevent future disruptions. Failure to do so may result in your account permanently being banned.

That is the bare minimum here.

11

Then no one should complain when they vibe-behaved their way into a ban :joy:

Virtually no one reads the law; the police are still going to arrest you without warning anyways.

This is no different.

2 Likes