Free AI API tiers in 2026: what each one actually limits

If you searched for a free AI API expecting to generate images or video without paying, Google's own pricing page has bad news. Every Nano Banana variant, all three Imagen 4 models, every Veo version and the Lyria music models list their Free Tier price as the same two words: "Not available"^[1]. The free tier is real, but it covers text.

That distinction is buried in a table most people skim past, and it explains a lot of frustration. Below is what each major provider's free tier actually permits, pulled from their own pricing pages in July 2026, plus the row in Google's table that almost nobody reads.

TL;DR

• Google's free tier has no image, video or music models. Nano Banana, Nano Banana 2, 2 Lite, Pro, Imagen 4, Veo 2, Veo 3, Veo 3.1 and Lyria 3 all show "Not available" under Free Tier^[1].
• Google marks free-tier input as used to improve its products. Every model block carries the row "Used to improve our products", answered Yes for Free Tier and No for Paid Tier^[1]. Google does not spell out what "improve" covers.
• Free-tier rate limits are no longer published. Google says limits "can be viewed in Google AI Studio" and that "Specified rate limits are not guaranteed and actual capacity may vary"^[2].
• OpenAI documents a free test request, but no ongoing allowance. Its quickstart congratulates you on "running a free test API request"^[8]; the pricing page lists no free inference allowance beyond that^[3].
• Replicate does have free limits; fal does not advertise any. Replicate lets you "run select models on Replicate for free, but after a bit you'll be asked to set up billing"^[7]. fal's pricing page carries no free-tier language at all^[4].
• Paying a fraction of a cent per image is a different problem from paying nothing, and it is a much easier one to solve.

What Google's free AI API tier actually covers

Google's Gemini API pricing page lists a Free Tier column next to a Paid Tier column for every model. Twenty models show "Free of charge" in that column: Gemini 3.6 Flash, 3.5 Flash, 3.5 Flash-Lite, 3.5 Live Translate, 3.1 Flash-Lite, 3.1 Flash Live Preview, 3.1 Flash TTS Preview, 3 Flash Preview, 2.5 Pro, 2.5 Flash, 2.5 Flash-Lite, 2.5 Flash Native Audio, 2.5 Flash Preview TTS, the Embedding models, Robotics-ER 1.6 Preview and Gemma 4^[1].

Every generative media model is on the other list. The table below is copied from that page.

Google is explicit about it in prose too. The Veo 3.1 description reads "available to developers on the paid tier of the Gemini API"^[1].

So when someone asks whether Nano Banana has a free API, the answer from Google's own documentation is no, on any of the four variants, for input or output. The Grounding with Google Search line on those models is also marked "Not available" for free tier^[1].

The row nobody reads

Scroll to the bottom of any model block on that pricing page and there is a row labelled "Used to improve our products". For Free Tier it says Yes. For Paid Tier it says No^[1].

That row appears under every model, including the text models that genuinely are free of charge. Google does not define "improve our products" on that page, so the honest reading is the narrow one: on the free tier your prompts, and whatever images or documents you attach, are retained for Google's own use in a way the paid tier excludes.

For a weekend project that is a fine trade. For anything touching client work, internal documents or user-submitted content, it is worth reading twice before you paste a production prompt into a free-tier key. The paid tier flips that row to No, which is the actual product difference between the two tiers on many models.

Rate limits you cannot look up before signing up

Until recently you could read Google's free-tier requests-per-minute and requests-per-day figures straight from the docs. That table is gone. The rate limits page now says limits "depend on a variety of factors (such as your usage tier) and can be viewed in Google AI Studio", followed by "Specified rate limits are not guaranteed and actual capacity may vary"^[2].

The usage-tier table is still there, and it is short. The Free tier qualification is "Active project or free trial" with no billing cap. Tier 1 requires you to "Set up and link an active billing account" and caps at $250. Tier 2 needs $100 paid plus three days, Tier 3 needs $1,000 paid plus thirty days^[2].

One free-tier ceiling is still published, and it is worth knowing: Search grounding on Gemini 2.5 Flash is "Free of charge, up to 500 RPD", with that daily limit shared with Flash-Lite^[1]. Everything else about free-tier throughput is per-project and only visible from inside your own account.

Two practical consequences. You cannot capacity-plan a free AI API before you have an account. And moving off the free tier is a billing-account step, not a plan upgrade, so the "just add a card later" migration is a little more involved than it sounds.

What the other providers mean by free

The picture outside Google is simpler, and mostly it means there is no free AI API to discuss at all.

OpenAI. The quickstart ends with "Congrats on running a free test API request!" and then points you at billing^[8]. OpenAI does not publish how many such requests a new key gets, only that the walkthrough is free to complete. Beyond that the pricing page lists per-token and per-image rates with no free inference allowance; its only free quantities are storage, 1 GB for file search and 1 GB per account per month for ChatKit uploads^[3]. Free ChatGPT is a consumer product and does not come with API credits.

fal.ai. The pricing page is titled Pay-Per-Use. Reading it end to end, there is no mention of a free tier, a trial or starter credits^[4].

Replicate. This is the one real free tier in the group, and it is deliberately vague: "You can run select models on Replicate for free, but after a bit you'll be asked to set up billing"^[7]. No model list, no quota, no reset window. Past that point the pricing page applies: "You only pay for what you use on Replicate. Some models are billed by hardware and time, others by input and output"^[5] — and hardware-billed models charge for occupied GPU time, so a slow model costs more than a fast one at identical output.

Notice the shape of what is on offer. A documented free walkthrough call, or an unspecified allowance on an unspecified subset. Nobody in this market is giving away image or video generation at scale, because each generation has a hard compute floor underneath it. Text is cheap enough to subsidise for goodwill. A 4K image is not.

What a fraction of a cent actually buys

Once you accept that generative media is not free anywhere, the question changes from "who is free" to "how small can the first bill be". That is a much better question, because the answer is now genuinely small.

Here are current per-generation rates on reAPI, where 1 credit is $0.001 and you are billed only for completed generations^[6]:

Text models bill per token rather than per generation, and this is where Google's free tier is a genuine competitor. The reason to pay for text is that data row, not the capability^[1]:

Two of the image rates are worth sitting with. Nano Banana 2 Lite at $0.02 per 1K image is below Google's own paid rate of $0.0336 for the same model^[1][6], and it does not require a linked billing account. And a low-tier GPT Image 2 render at half a cent means testing a prompt twenty times costs a dime.

Signing up on reAPI includes starter credits with no card required, which is enough to work through the quickstart and see real output before you decide anything. Per-model rates for everything not listed above are on the pricing page, alongside each model's official list price for comparison.

One more constraint that has nothing to do with price: Google embeds a SynthID watermark in image output from its models. If your use case involves resale or downstream editing, read up on what SynthID does to Nano Banana output before you commit to a pipeline.

FAQ

Can I get an AI API for free?

For text, yes. Google's Gemini API offers around twenty text, audio and embedding models at no charge, including Gemini 2.5 Pro and the 3.x Flash family^[1]. For image, video or music generation, no major provider offers a free tier as of July 2026.

Does Nano Banana have a free API?

No. All four Nano Banana variants list Free Tier as "Not available" for input, output and search grounding. Paid rates run from $0.0336 per 1K image on 2 Lite to $0.24 per 4K image on Pro^[1].

Is there a free Veo 3.1 API?

No. Google describes Veo 3.1 as "available to developers on the paid tier of the Gemini API" and lists Free Tier as "Not available" for the Standard, Fast and Lite variants^[1].

Does Google train on my free-tier API requests?

Google's wording is "Used to improve our products", not "train", and the page does not define the term. That row reads Yes on Free Tier and No on Paid Tier for every model^[1]. It applies to the models that are free of charge, since the media models have no free tier at all.

What are the Gemini free-tier rate limits?

Per-model request limits are no longer published. The docs direct you to view your own limits in Google AI Studio and add that "Specified rate limits are not guaranteed and actual capacity may vary"^[2]. The one exception still printed is Search grounding at 500 requests per day on the free tier^[1]. Anyone quoting you a specific free-tier RPM figure for a model is reading a cached version of that page.

Do OpenAI API keys come with free credits?

The quickstart explicitly describes its walkthrough call as "a free test API request"^[8], but OpenAI does not document a standing free allowance. The pricing page lists none; its only free quantities are 1 GB of file-search storage and 1 GB per month of ChatKit upload storage^[3].

Is Replicate free to start?

Yes, within limits Replicate does not quantify: "You can run select models on Replicate for free, but after a bit you'll be asked to set up billing"^[7]. Which models and how much are both unstated, so it is a way to try something rather than something to plan against. After that, "You only pay for what you use"^[5].

What is the cheapest way to test an image model?

Per-image rates are now low enough that testing is a rounding error. A Stable Low render on GPT Image 2 is $0.005, so twenty test images cost ten cents^[6]. That is a more predictable path than working around free-tier quotas you cannot see in advance.

Picking a tier without getting stuck

If your workload is text, take Google's free tier. It is genuinely free and it covers capable models. The cost that is not measured in dollars is the "Used to improve our products" row, which matters for some projects and not others.

If your workload is images, video or music, skip the search for a free AI API and optimise for the smallest possible first bill instead. That means per-generation billing rather than subscriptions, no billing-account prerequisite, and rates you can read before you sign up. A free AI API that excludes the models you came for is not a starting point. Half a cent per image is.

References

1. Google. Gemini Developer API pricing — Free Tier and Paid Tier rates by model. Retrieved July 2026 from ai.google.dev/gemini-api/docs/pricing
2. Google. Gemini API rate limits — usage tiers and tier qualification. Retrieved July 2026 from ai.google.dev/gemini-api/docs/rate-limits
3. OpenAI. API pricing. Retrieved July 2026 from platform.openai.com/docs/pricing
4. fal.ai. GenAI API pricing — pay-per-use. Retrieved July 2026 from fal.ai/pricing
5. Replicate. Pricing. Retrieved July 2026 from replicate.com/pricing
6. reAPI. Model pricing — GPT Image 2, Nano Banana 2 Lite, Seedance 2.0 Mini. Retrieved July 2026 from reapi.ai/models
7. Replicate. Billing — free limits. Retrieved July 2026 from replicate.com/docs/topics/billing
8. OpenAI. API quickstart. Retrieved July 2026 from developers.openai.com/api/docs/quickstart

Further reading

• reAPI. Nano Banana API free tier. reapi.ai/blog/nano-banana-api-free-tier
• reAPI. What SynthID does to Nano Banana output. reapi.ai/blog/nano-banana-watermark-synthid