Gemini Free Tier: What 20 API Requests a Day Really Gets You

The Gemini free tier comes in two forms: the Gemini app is free with usage limits, and the Gemini API free tier gives you about 20 requests a day per model on Flash models. Those daily API counts reset at midnight Pacific time, and Google tracks them per project, not per API key.

I learned what that means in practice in October 2026, when I wired Gemini's free API into the newsletter tools inside the Boredlisted dashboard. Its job is small: suggest email subject lines for each issue. Below is what the cap, the speed and the error rate looked like.

If you're still deciding whether any AI plan is worth money, my guide to which AI subscription is worth paying for covers the consumer plans. This article stays on the free tier and the developer side of Gemini.

Gemini app vs Gemini API: two different free tiers

Most confusion about the Gemini free tier comes from mixing up two products. The Gemini app is the chatbot on the web and on your phone. The Gemini API is how software, like my newsletter tool, talks to the same models through a key you create in Google AI Studio, Google's workspace for developers.

If you're new to the assistant itself, start with how Gemini works and what it costs. Neither product below requires a credit card to start.

Google moved the app to those compute-based limits on May 17, 2026, and paid Google AI plans raise them. If you mostly want help inside Google Workspace, my prompts to use Gemini in Docs and Gmail are the better fit, and you won't need an API key at all.

What are the Gemini API free tier limits right now?

Is Gemini API free? Yes, on selected models, with no card required. Google's Gemini API pricing page describes the free tier as limited access to certain models, with free input and output tokens.

That table reflects Google's pricing page on October 20, 2026. Model names rotate fast, so check the page before you hard-code an ID.

The Gemini API rate limits work on three meters, and going over any one of them triggers an error:

• Requests per minute (RPM): how many calls you can make to a model in any one minute.
• Tokens per minute (TPM): how much input you send per minute. A token is a small chunk of a word.
• Requests per day (RPD): your daily allowance per model. For a small tool, this is usually the one that bites.

Two details in Google's rate limits documentation matter more than the numbers. Limits apply per project, not per API key, so a second key in the same project gets you nothing. And the daily count resets at midnight Pacific time.

That's 3 a.m. Eastern in the US and 8 a.m. in the UK for most of the year. Here in Manila, it lands in the middle of the afternoon.

Here's the catch: Google no longer prints free-tier numbers in its docs. They appear in the Rate Limit view in Google AI Studio, for your own project, and Google says they aren't guaranteed. On my project in October 2026, each Flash model showed 20 requests a day. Treat that as a dated reading, not a promise.

What does 20 requests a day get you in a real workflow?

While I built and tested the newsletter tools in my dashboard, every click on a generate button sent one request to Gemini.

That testing alone used up the day's Gemini 2.5 Flash allowance. After that, the two tools that call Gemini ran on their backup models until the counter reset at midnight Pacific. Nothing broke, but the output changed, which I'll get to below.

Each model has its own counter

The 20-request cap is per model, so when 2.5 Flash ran out, a different Gemini model in the same project still had its own full allowance. A fallback to a second free model isn't only a reliability trick. It adds a whole extra day's quota.

Do the math before you build. If a tool makes one call per newsletter issue, 20 a day is plenty for a daily send. If you regenerate five times per issue while tuning the tone, four issues use up the whole day.

Gemini 3 Flash vs 2.5 Flash: is the slower model worth it?

Speed was the second surprise. Gemini 3 Flash wrote the best subject lines in my tests, but each answer took 10 to 15 seconds. Gemini 2.5 Flash came back in about 5 seconds.

For a button I click while writing, 5 seconds feels fine and 15 feels broken. Google's troubleshooting guide notes that Gemini 3 models think by default, which adds latency, and that you can lower the thinking level when speed matters more than depth. That's the first setting I'd test before giving up on the better model.

A tighter prompt is the other lever. My library of copy-ready AI prompts is a good place to borrow a structure from.

How reliable is the Gemini free tier?

Not very, at least during my October tests. The free models often came back busy instead of returning subject lines. Google's wording is usually about the model "experiencing high demand," and it arrives as a 503 (unavailable) error. That's Google's servers, not your quota.

My fallback chain handled it. When one model was busy, the request moved to the next one, and the newsletter tools kept working. The worst case was a single click that fell all the way through to OpenAI. From click to subject lines, it took 29 seconds.

That's the hidden cost of a deep chain: every failed attempt adds its own wait before the next model gets a turn. And when a Flash-Lite backup answered, the subject lines were noticeably plainer. The tool stayed up. The quality didn't.

Who should use the Gemini API free tier, and when should you pay?

Here's where I've landed after the test:

That last row deserves its own note. Under Google's Gemini API additional terms, free-tier prompts and responses can be used to improve Google's products, and human reviewers may read them. Public article titles are fine to send. Subscriber lists and client drafts are not.

UK readers get a different deal. In the UK, the European Economic Area (EEA) and Switzerland, Google applies its paid-tier data rules even to free usage. It also requires paid services when you offer an app to users there. US developers have no such rule, so check where your users are, not just where you are.

Turning on billing moves your project from the Free tier to Tier 1, which Google says usually happens instantly. That's separate from Google AI Pro, the consumer plan I cover in my comparison of the $20 AI plans. A Google AI subscription doesn't raise your API limits.

How do you build your own fallback chain?

A fallback chain is an ordered list of models your code tries, one at a time, until one answers. Mine follows this shape:

1. Primary model: the one whose output you like best, as long as you can live with its speed.
2. A second Gemini model: a different Flash model, which has its own daily counter.
3. A Flash-Lite model: a lighter safety net. Expect plainer writing.
4. Another provider: OpenAI in my case, so a busy Google day doesn't stop the tool.

A few rules keep the chain well behaved:

• Move to the next model on a 429 (quota or rate limit) or a 503 (busy). Google lists both as errors worth retrying.
• Stop on 400 and 403 errors. They point to a bad request or key, and retrying won't fix them.
• Give each attempt a time limit, so one slow model can't stall the whole chain.
• Record which model answered and show it next to the result, so you know when to rewrite.
• Recheck your model IDs regularly. Google retires and renames models often, and an old ID fails with a not-found error.

If you'd rather not depend on any cloud quota for the last step, you can run an open source assistant on your own hardware and point the final fallback at it.

How do you fix Gemini 429 and busy errors?

Google recommends exponential backoff: wait about 1 second, then 2, then 4, then 8, with a little random jitter so many clients don't retry at once. Its official SDKs already do this for 429 and 5xx errors, so check yours before writing your own.

One fix that doesn't work: creating a new API key. Limits apply per project, so a fresh key in the same project starts with the same empty tank.