Most guides to this answer the wrong question. They rank free tiers by quota, which has two problems in late 2026. Google, Groq and Mistral have all moved their per-model free limits behind a logged-in console, so any article quoting exact free Gemini numbers is citing a page that no longer exists. And quota was never the expensive part.
The expensive part is who reads your prompts. That varies enormously between providers offering superficially identical free tiers, it is documented in every case, and almost nobody checks. Here is what the primary sources said on 22 September 2026.
Who trains on your free-tier traffic
| Provider | Trains on your data? |
|---|---|
| Cloudflare Workers AI | No, at any tier, without explicit consent |
| Groq | No. Contractually barred unless you instruct otherwise |
| Scaleway | No. Also not readable by them or the model vendors |
| xAI API | No, without explicit permission |
| OpenAI API | No, unless you opt in |
| Alibaba Model Studio | No, though calls are stored |
| SambaNova | Not your content. Query logs, yes |
| Z.ai | API no. Consumer chat yes, under a perpetual licence |
| OpenRouter | Per-account toggle, separate for free and paid models |
| Mistral | Yes by default. Opt-out toggle in the admin console |
| Nebius | Yes by default, for speculative-decoding draft models |
| Google Gemini free tier | Yes, and human reviewers may read it |
| Cohere trial keys | Yes, and trial users cannot opt out |
| NVIDIA NIM | Yes, inputs and outputs. No opt-out offered |
| Black Forest Labs | Yes, inputs and outputs |
Google is the one worth pausing on, because it is also the most generous. Its pricing table carries a column reading “Used to improve our products”, and every free-tier row says Yes. The terms add that human reviewers may read, annotate and process your input and output, after disconnecting it from your account and key. Move to the paid tier and that stops entirely. The free tier is not a loss leader, it is a data trade, and Google is unusually honest about saying so in a table.
Cohere is the sharper edge. Trial keys are trained on and the opt-out is an enterprise feature, so the free user has no route to decline.
Genuinely free, no card, no expiry
These four are the real ones.
| Provider | What you get | Card |
|---|---|---|
| Google AI Studio | Every current Gemini text model free of charge | No |
| Groq | Permanent free plan, downgrade back any time | No |
| Cloudflare Workers AI | 10,000 Neurons per day, resets 00:00 UTC | No |
| Mistral | Permanent free mode with included monthly usage | No |
Cloudflare deserves more attention than it gets. Ten thousand neurons a day is modest but it genuinely recurs, needs no card, and comes with the cleanest data position of any provider here. For text generation the cap is 300 requests a minute.
Two more with published numbers, which is rarer than it should be. OpenRouter gives its :free model variants 20 requests a minute and 50 a day, rising to 1,000 a day once you have bought $10 of credit at any point. Cohere trial keys allow 20 chat calls a minute and 1,000 API calls a month.
Hugging Face credits every free account $0.10 a month, which sounds like a rounding error and is, but it is real, recurring, and routes to genuine image and video models.
Why nobody can quote you a quota any more
Through 2025 these articles were easy to write, because providers published a grid of requests per minute and tokens per day. During 2026 the three most-cited free tiers stopped doing that.
Google now says rate limits depend on your usage tier and can be viewed in AI Studio. Groq says the current exact limits for your organisation are on the limits page in your account settings. Mistral puts them in the admin console. In all three cases the number is real, it is simply account-specific and no longer public.
This has a practical consequence. If a roundup quotes you a precise free Gemini or Groq figure today, it is either stale or invented, because the source page was withdrawn. The only reliable answer for your own account is to log in and read it, which is unsatisfying to write and accurate.
It also means the durable way to compare free tiers is by the things providers still publish: whether a card is required, whether the tier expires, and what happens to your data. Those are contractual and they change slowly.
Free with a catch
NVIDIA is the one to read carefully. The terms are titled “API Trial Terms of Service” and they prohibit production use outright, including use of generated content in production. They also state NVIDIA collects your inputs and outputs to improve its AI models, with no opt-out anywhere in the document. Free, they keep it, and you may not ship on it.
Cerebras advertises free access but the docs gate $5 of credits behind a verified payment method, and they expire after 30 days.
Nebius gives $1 valid for 30 days and requires a bank card before onboarding completes.
ElevenLabs offers 10,000 credits a month, and then the legal terms remove most of the value: the free plan carries no commercial licence, and anything you publish from it must credit ElevenLabs in the title.
Not free, or no longer alive
GitHub Models was fully retired on 30 July 2026. The playground, catalog, inference API and bring-your-own-key are all gone. It still appears on most roundups of this topic.
OpenAI, Anthropic, Together and DeepSeek have no free API tier. Anthropic grants new users a small unpublished trial credit. Together requires a $5 minimum purchase. Fireworks gives $1 and then needs a card.
Media generation, where the myth lives
There is no free Nano Banana API key. Google’s pricing page reads “Not available” on the free tier for every image model, for Veo 3.1 in all three variants, and for its music models, while still reporting Yes under “used to improve our products”. Articles promising otherwise are describing the consumer app, not the API.
What is actually free for media:
- Cloudflare Workers AI, the strongest option. FLUX-1-schnell costs 4.80 neurons per 512 by 512 tile against a 10,000 neuron daily budget. Other image models burn it far faster, with some Leonardo models above 500 neurons per tile.
- Replicate offers free runs on a curated collection including Imagen 4 and FLUX variants, explicitly without a credit card. The number of runs is not published.
- Google TTS, where two Gemini Flash preview voices are free of charge.
Note also that every Google-generated image carries a SynthID watermark, paid tiers included.
One provider is genuinely free forever with no account at all. Pollinations.ai serves images anonymously at one request every 15 seconds, dropping to one every five if you register. Free output is watermarked unless you have an account, its terms cover the non-commercial site only, and commercial use depends on each underlying model’s own licence. Within those limits it is the only no-signup, no-card, no-expiry image API I found.
The separate-wallet trap
Three image providers run two currencies, and their free consumer credits never reach the API. Leonardo, Recraft and Ideogram all state this in their own docs, and all three are routinely listed as free API options by people who tried the web app.
| Provider | Consumer free tier | API |
|---|---|---|
| Leonardo | Daily free tokens | Prepaid only, separate account |
| Recraft | Free plan credits | Separate API Units, $1 = 1,000 |
| Ideogram | Weekly free credits | Prepaid, positive balance required |
Recraft carries the sharpest trap in this entire survey. Images generated on its Free plan are publicly visible and remain the property of Recraft, with no commercial use permitted. Free-plan content is also used for training with no opt-out, while API traffic never is. The free tier and the paid API have opposite data and ownership terms.
Ideogram, meanwhile, requires a “Powered by Ideogram” credit in any application built on it, the same in-interface attribution pattern Runway uses.
Video is bleaker still. Across Luma, Runway, Kling, Pika, MiniMax and Higgsfield, not one offers a free-forever API tier. Pika states it outright in its own docs: there is no free generation tier and no separate test environment. Runway requires a $10 minimum before your first call, Higgsfield $5. The only confirmed free starting quota on a video API belongs to Alibaba Model Studio, and it is a one-time new-user grant, valid 90 days whether you use it or not, available only in the Singapore region.
Two traps in that group are worth naming because both are easy to miss and both survive into production.
Runway requires attribution in your interface. Not a watermark on the file, a “Powered by Runway” credit with a link displayed in your own UI. For an API integrator that is the functional equivalent of a watermark, and it is a design constraint rather than a licensing footnote.
Luma’s free output may not be used commercially. Its terms license commercial use of outputs produced under a paid subscription that allows it, which excludes anything you made while evaluating. Free users also grant Luma the broadest training and display licence of anyone checked here.
One genuinely good norm across this whole category: every provider that documents the question, including Luma, Kling, Runway, Higgsfield and Alibaba, does not charge for failed generations.
When free runs out, what is actually cheap
Free tiers are for evaluation. The moment something works, the question becomes cost per unit, and the spread here is wider than most people expect.
For text, DeepSeek is the aggressive option. Its flash model runs $0.15 per million input tokens on a cache miss and $0.60 per million output tokens at off-peak rates, and off-peak is half price covering most of the week. Peak is only 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so the discount is closer to the default than the exception.
For images, the cheapest verified per-unit prices come from providers nobody lists in these articles. Together’s Juggernaut Lightning Flux is $0.0017 per megapixel. Black Forest Labs starts at $0.014 an image for the smallest FLUX.2 variant, though note its terms state it may use inputs and outputs to train its models, which puts it in the same category as the free tiers above rather than above them.
For video, xAI publishes five cents a second, so a 15-second clip is under a dollar. That number matters because it sets the ceiling on what an aggregator’s discount can be worth. If direct access costs a dollar, a middle layer saving you thirty cents has to be worth the dependency it adds.
One caveat on this section. Several budget providers frequently named in cheapest-API roundups were not verified here, and I would rather leave them out than repeat a figure I have not read on the provider’s own page.
What I would actually pick
For text, start with Groq or Cloudflare. Both are permanently free, neither wants a card, and both are contractually barred from training on you. Add Gemini’s free tier when you want the strongest model available at no cost and the work is not sensitive, remembering that the trade is human-reviewable data and that callers serving users in the EEA, Switzerland or the UK are required to use the paid tier.
For images, Cloudflare is the only recurring free option worth building on.
And if you are wiring several of these into agents, the thing that bites later is not quota, it is that each provider has a different key, a different rate limit and a different data policy, and that state ends up spread across whichever tools happen to hold it. Sharkly works on that layer, around the tools doing the execution, with model usage running on the subscriptions and keys you configure in them rather than on anything it resells.
