The purpose of this post is not to simply complain, but to deconstruct our current situation from the perspectives of software engineering and business logic. The conclusion is regrettable: this is not a bug that can be fixed in the short term. I strongly recommend that everyone pause their renewals until Google provides a substantive, architectural-level solution.
-
The Invisible Compute Black Hole and “Credits”
Many people complain: “I only asked few questions, why my quota gone?”
This is not a bug, but the unfortunate result of a collision between Antigravity’s underlying Agentic architecture and an extremely opaque billing system. When you issue a simple refactoring command, model is actually performing over a dozen self-iterations (looping) and sandbox tests in the backend, each involving massive context (often hundreds of thousands of tokens).
Google discovered that the cost of this 1M Context Agent loop is terrifyingly high, so the finance department hit the brakes. However, instead of transparently disclosing token consumption, they introduced an extremely opaque “AI Credits” system.
As developers, we are accustomed to a clear Pay-as-you-go (token-based) model. But now, Google is using consumer-grade “blind box credits” to package enterprise-grade compute consumption. You don’t know how many credits 1 million tokens equal; you are only guided to purchase expensive Top-up Credits when your quota runs out. This is essentially a business isolation tactic designed to drive away high-consumption, power users. -
More Bugs with Every Fix
Antigravity is now a terrifyingly cobbled-together vehicle:
- The core model comes from DeepMind and Anthropic.
- Subscription and billing are tied to Google One (consumer division).
- Compute quotas and settlement go through Google Cloud (enterprise division).
- Account security (WAF firewall) is yet another independent system.
When you encounter a bug where a “paid account is downgraded to Free Tier,” frontline customer service has no authority to handle it. Engineers apply patches to fix quota issues, but they may inadvertently break the Cloud billing state, triggering new bugs like context shrinking to 20k. In this quagmire of cross-departmental high coupling and unclear accountability, “patching” only accelerates the system’s collapse. The only truly effective solution is a “rewrite from scratch,” but under Google’s promotion-driven development culture and pressure for quarterly earnings, no one is willing or has the time to take on such a thankless refactoring job.
-
Fatal Risk: Indiscriminate WAF Bans
What is currently most unacceptable is the overly sensitive anti-abuse mechanism (WAF). Because system bugs cause the client to constantly retry requests, Google’s security system mechanically flags these legitimate developers as “DDoS or malicious API abuse,” immediately imposing a 7-day lockout.
Until Google completely reconstructs this billing and permissions system and makes billing 100% transparent, I strongly recommend that you take the following action: immediately cancel your AI Ultra subscription, switch to a transparent API model, or embrace alternatives like CC or Codex.
A developer’s time and energy should be spent on creating great products, not on staring at forums every day, waiting bitterly for a billing bug to be fixed.