To everyone who has been stuck for days with “Quota limits exceeded” in Build mode despite a paid key and a dashboard that looks almost empty: I took this apart in my own project. Here is what I found.
What happened
- Build mode, Gemini 3.8 Flash, paid API key, project on Tier 1.
- Three consecutive agent runs (about 110, 48 and 36 seconds) edited files, and then each one failed with the same error.
- At the same time, the UI said “using Gemini without limits.”
What the Rate Limit dashboard showed at the time of failure
- RPM: 11 / 1,000. Fine.
- RPD: 54 / 10,000. Fine.
- TPM: 2.44M / 2M. Exceeded.
- Error dashboard: HTTP 429 TooManyRequests.
- No spend cap, and the monthly bill was negligible.
So this is not a billing problem, not a wrong key, and not a daily limit. It is the tokens-per-minute limit.
Why it happens
- According to Google’s documentation, rate limits apply to the Google Cloud project, not to an individual API key. Every key in the same project shares the same quota, so switching keys does not help.
- The Build agent does not compact the conversation. On every turn it resends a large part of the history, the code, the tool traces and the reports, sometimes with several model calls in the same minute. In a long working session that is enough to pass 2M tokens per minute even when you have sent almost no requests.
- Pressing Retry immediately resends the same heavy payload, so the limit is hit again.
- Worst of all: the agent edits files and installs packages before it fails, leaving your code in a partial state.
How to check whether this is your case
- Open your project’s Rate Limit dashboard in AI Studio right after the failure and look at TPM, not only RPM and RPD.
- Check the error dashboard for 429.
- If your TPM really is far below the limit (there are reports here of 0/60 Antigravity usage), your case may be a different Google-side issue, and it is worth posting your numbers.
What worked for me as a workaround
- Do not press Retry again in the same chat.
- Open a New chat inside the same app. The code stays, and the huge conversation payload is reset.
- Wait 60 to 90 seconds after an error so the one-minute window clears.
- Give the agent one focused task at a time, with no full project scans and no giant handoff reports.
- Check your project’s dashboard for a model with a higher TPM limit. In my project, the Flash Lite models get 4M TPM on Tier 1, compared with 2M for Gemini 3.8 Flash.
- If your app itself also calls Gemini, move its server key to a separate Cloud project so it does not eat the agent’s quota.
- After every failure, check which files the agent already changed before you continue.
What Google must do
- Acknowledge the problem publicly and give a date for a fix.
- Show in the error message which limit was exceeded and when it is safe to resume.
- Compact the agent’s context instead of resending the whole history.
- Apply code changes atomically, or roll them back when a run fails.
- Remove “without limits” from the UI.
And now, without the polite wording
Google is one of the largest companies in the world. Its products serve billions of users, and it rents out its TPU capacity to other companies. Yet since at least August 24, when the first Tier 2 report of this failure was posted here, it has been unable to give its paying developers a coding agent that does not choke itself. Since then, at least ten threads about this same failure have been opened on this forum, and the only responses from Google staff have been “I have DM’ed you” and a question about whether the user is on the free or paid tier. No status notice, no explanation, no timeline.
A paid development environment with a 2-million-tokens-per-minute ceiling that Google’s own agent blows through in a single session is a joke. You cannot try to take over the entire business market while abandoning the developers who are already paying you. We are not asking for a favor. We are asking for the product we pay for to work, and for someone at Google to say publicly what is going on.
What this cost me
This has been the hardest month of my work. I am a solo developer, about to deploy an MVP, and instead of shipping I have spent my time fighting an opaque error, auditing half-finished edits the agent left behind, and waiting for a quota window to clear. The delay has hit my income directly.
This is how a platform hurts the weakest users without even noticing. A large team can absorb a lost month. A solo developer or a small startup cannot. We are exactly the builders AI Studio is marketed to, and we are the ones left paying for a failure that Google has not even acknowledged. That is a shame, and it does not have to be this way.
Red flags, and why I now pay OpenAI to code inside AI Studio
Look at the full picture: a coding agent that cannot manage its own context, a quota error that hides which limit was hit, file edits applied before the run is known to succeed, a UI that says “without limits” while enforcing a hard ceiling, and more than a month of reports answered with private messages. Every one of these is a red flag. Together they describe a product that was shipped to paying developers before it was ready.
Gemini 3.8 Flash was launched with a lot of pride. In real, sustained development work, it is the weakest link in my stack. The Build agent is so weak that I ended up paying OpenAI for Codex, an autonomous coding agent, to write the code inside AI Studio instead of Google’s own agent. Think about that: a Google development environment, where the most reliable way to get work done is to bring in a competitor’s agent. Over this period, the Build agent burned hundreds of hours of my development time, which by my own estimate is worth around $250,000. Codex did the same work for about $100.
My recommendation: how to stop burning money
If you rely on Gemini Flash for 90% of your coding in AI Studio, this is what worked for me instead:
- Download a copy of your AI Studio project to your computer.
- Let Codex read the project and learn the codebase first. Tell it explicitly not to build anything at this stage.
- Then use Codex through its Chrome extension to write the code changes directly in the AI Studio environment, instead of the Gemini Flash Build agent.
- Keep Gemini for what it does well, and stop paying, in time and in money, for an agent that chokes on its own context.
A word to Google
Use common sense. Serve your own users first, not the competing AI companies you sell hardware to. Right now it looks like the race for money and power comes first, while the developers who pay you and build on your platform absorb the damage. And the real cost goes beyond us: you are hurting the global project that AI actually is, building “the brain”, by pushing away the very people who are building it with you.
And still: Google gave the world some of the most important foundations of modern AI. Gemini has real potential, and many of us chose AI Studio because we believed in it and wanted it to succeed. That is exactly why this hurts.
The moral is simple: a platform is only as strong as the trust of the developers who build on it. Earn that trust back with a public answer and a real fix, and we will be the first to say so here.
If you have hit this too, post your TPM, RPM and RPD numbers from the moment of failure. The more data is public, the harder it will be for Google to keep answering in private messages.