The Gemini API free tier still works for text models, but image generation and Veo video now require billing. This article breaks down per-project rate limits, which models stay free, what each paid tier requires, and where to make AI images without a Google Cloud account.
The Gemini API free tier still exists, but it no longer looks like the one people described a year ago. Text models can be called without a credit card, while every model that draws pictures or renders video now sits behind billing, and the daily quotas on the remaining models have shifted more than once. If you are budgeting a side project, a prototype or a classroom demo around these limits, the details decide whether your app works on launch day or throws errors by lunchtime.
This article lays out what the free tier includes, which models you can call without billing, how rate limits behave in a real project, and why image generation is the sticking point. You will also see what the paid tiers require and where to make AI images without a Google Cloud account.
💡 Quick answer: Text models such as Gemini 2.5 Flash have free access with tight per-project quotas. Image models like Nano Banana and Imagen 4, plus Veo 3.1 video, need a billed project.
Every new Gemini API project starts on the free tier. The only requirement is an active project, so there is no billing account to link and no card to enter. That makes it the fastest way to test a prompt, wire up a prototype, or check how a model handles your own data.
Google measures usage per project, not per API credential. Creating five credentials inside one project does not give you five quotas, because they all draw from the same pool. Splitting one app across several projects to stack allowances is fragile, and it is worth avoiding. Plan around a single project's allowance and design your app to live inside it.
Four counters decide whether a request goes through:
Exceeding any single counter triggers a rate limit error, even when the other three look healthy. A chatbot that sends short messages usually trips RPM first. A document summarizer that sends long files trips TPM. A hobby app that runs all afternoon trips RPD.
| Limit | What it counts | Who hits it first |
|---|---|---|
| RPM | Requests in a 60 second window | Chat apps, bursty scripts |
| TPM | Input tokens in a 60 second window | Long documents, big contexts |
| RPD | Requests since the last reset | Batch jobs, all day tools |
| IPM | Images per minute | Image generation loops |
Google's documentation no longer prints a per-model free tier table. The exact RPM, TPM and RPD for your project appear in the rate limit view inside Google AI Studio, and they can differ from project to project. In December 2025, developers reported steep cuts, with Gemini 2.5 Flash falling to roughly 20 requests per day in some reports. Third-party tables from early 2026 listed ranges of 5 to 15 RPM, around 250,000 TPM and 100 to 1,000 RPD. Treat those as reported figures, not promises.
💡 Tip: Read limits from your own AI Studio dashboard and from the error messages your app receives. Hardcoding a number from a blog post, including this one, is how apps break after a quiet quota change.
Not every model shares the same free status. At the time of writing, the pricing page draws a clean line between text models and everything that produces media.
The Flash family is the free tier's home. According to the pricing page, the Gemini 3.x Flash models and Gemini 2.5 Flash list input and output as "free of charge" in the Standard tier. The lighter Flash-Lite variants go further and stay free across Standard, Batch and Flex. Flash-class text-to-speech models are also free in the Standard and Priority tiers.
Gemini 3.5 Flash is the model most newcomers should start with, since it handles chat, code and image inputs at a speed that suits quotas this small. Batch and Flex do not come free on the larger Flash models, so a bulk job on those tiers means billing.
Everything that outputs pictures, music or video needs a paid project. Pro-class text models are the gray area, so confirm them in AI Studio before building on them.
| Model family | Free tier | Notes |
|---|---|---|
| Gemini Flash and Flash-Lite text | Yes | Tight per-project quotas |
| Gemini 3.1 Pro Preview | Not in third-party checks | Verify in AI Studio |
| Nano Banana 2 (Gemini 3.1 Flash Image) | No | Billing required |
| Nano Banana Pro (Gemini 3 Pro Image) | No | Billing required |
| Imagen 4 family | No | Billed accounts only |
| Veo 3.1 | No | Not available on free |
| Lyria music | No | Billing required |
A team choosing a model should ask one question first: does the feature need media output? If yes, the free tier is already off the table, and the real decision is whether to pay Google directly or use another route.
A limit on paper and a limit in production feel very different. On paper, 15 requests per minute sounds fine. In production, a page that fires four calls on every load will hit that ceiling with a handful of visitors.
A quick sizing exercise helps. Divide your daily request limit by the number of calls one user action makes. Suppose your project allows 100 requests per day and each action makes three calls: you can serve about 33 actions before the reset, which is a few dozen visitors at most. Run that math with the numbers shown in your own AI Studio dashboard before you promise anything to users.
When a quota trips, the API answers with HTTP 429. The message names the limit that failed, so read it before guessing.
| What you see | Likely cause | Fix |
|---|---|---|
| Errors in short bursts, then recovery | RPM | Queue requests and space them out |
| Errors only on long prompts | TPM | Trim context or split documents |
| Fine all morning, failing by evening | RPD | Cache results and cut call volume |
| Image model returns a zero quota | No free tier | Link billing or use another platform |
Exponential backoff with a little randomness handles the per-minute limits well. This pattern fits in a dozen lines:
import random, time
def call_with_backoff(send, max_tries=5):
for attempt in range(max_tries):
try:
return send()
except RateLimitError:
wait = (2 ** attempt) + random.random()
time.sleep(wait)
raise RuntimeError("Rate limited after all retries")
Backoff solves RPM. It cannot solve RPD, because that quota returns only at midnight Pacific time, no matter how politely you retry.
Four habits stretch a small free quota further:
This is where the confusion starts, because older advice still fills the search results.
Early in 2026, several posts reported that Gemini 2.5 Flash Image allowed up to 500 free images per day at 1024 by 1024. That claim is outdated. The current pricing page lists no free tier for any image generation model:
Then there is the ghost quota. In AI Studio, a free project can show a counter such as "0 / 25" next to Imagen, which looks like a daily allowance. A Google developer forum thread from February 2026 reports that requests still fail with a message saying the Imagen API is only accessible to billed users. Enabling billing does not add a free daily allowance either. The only free usage comes from the $300 introductory credit given to new Google Cloud users, valid for 90 days. After it expires, pay-as-you-go rates apply immediately.
💡 Watch out: A quota number on screen is not a promise of access. If a model sits in the "No" column of the pricing page, no amount of waiting will make it work on the free tier.
Video follows the same rule. Veo 3.1 is listed as "not available" on the free tier across its standard pricing variants, so a demo that renders clips needs a paid project from day one. If you want video clips without any API setup, PicassoIA hosts Veo 3.1, Veo 3.1 Fast and Veo 3.1 Lite for direct use.
Paid tiers raise the ceiling and open up the media models. Moving up happens automatically once you meet the conditions below.
| Tier | How you qualify | Spend cap |
|---|---|---|
| Free | Active project or free trial | Not applicable |
| Tier 1 | Billing account linked | $250 |
| Tier 2 | $100+ spent and 3+ days since first payment | $2,000 |
| Tier 3 | $1,000+ spent and 30+ days since first payment | $20,000 to $100,000+ |
Each tier also has a spend-based limit measured over a rolling 10 minute window: $10 for Tier 1, $50 for Tier 2 and $200 for Tier 3. A move from Free to Tier 1 takes effect instantly after billing setup. Later upgrades arrive within about 10 minutes of meeting the criteria.
Stay on the free tier if:
Move to Tier 1 if:
If your goal is pictures rather than text, there is a shorter path than a Google Cloud project. PicassoIA hosts Google's image and video models in the browser, so there is no Google Cloud billing account, no project setup and no quota dashboard to watch.
The Google lineup that needs billing through the API is available directly on the platform:
| Need | Model |
|---|---|
| Fast, flexible image creation and editing | Nano Banana 2 |
| High resolution detail | Nano Banana Pro |
| Lighter, quicker drafts | Nano Banana 2 Lite |
| Photorealistic stills | Imagen 4 and Imagen 4 Ultra |
| Speed over polish | Imagen 4 Fast |
| Chat, code and writing | Gemini 3.5 Flash and Gemini 3 Flash |
Because these are separate models, you can send the same prompt to two of them and compare the results. Imagen 4 suits photorealistic stills, while Nano Banana 2 is the better pick when you want to edit or fuse existing images. Picking by task rather than by name keeps your results consistent.
Nano Banana 2 is the model to try first, because it creates images from text and also edits or fuses existing ones.
Specific prompts win. A vague request like "a desk" gives a generic result, while "a developer's oak desk at golden hour, steam rising from a mug, 35mm lens, shallow depth of field" gives you something you can publish.
💡 Tip: Change only one thing per attempt. If the lighting is wrong, edit the lighting words and leave the rest alone. You will see exactly what each phrase does.
Partly. Text models in the Flash and Flash-Lite families have a free tier with per-project limits. Pro-class text models, every image model, and Veo video need a billed project.
No. The current pricing page lists no free tier for Nano Banana, Gemini image models or Imagen. Posts claiming hundreds of free daily images describe an earlier setup. For image creation without a Google Cloud project, use a platform like PicassoIA with Nano Banana 2 or Imagen 4.
Requests per day reset at midnight Pacific time. Per-minute limits recover as the rolling window passes, which is why backoff fixes bursts but not a spent daily quota.
The free tier is a fine place to test text prompts, but the moment you need a picture, a billing account becomes the price of entry on Google's side. You can skip that step. Open Nano Banana 2 on PicassoIA, type the scene you have in mind, and see the result in seconds. Try Imagen 4 when you want a photorealistic still, or Veo 3.1 Fast when a still is not enough. Run the same prompt through two models, compare the results side by side, and keep the one that fits your project. Every model is listed at picassoia.com/en/all-models, so you can compare your options before you commit to one.
