Free tiers are real, and most people asking "cheap API" would be better served by one of them — until a specific limit starts costing them more time than a paid balance would cost money. This page is about finding that point.
Every free tier is bounded by at least one of four things. Which one binds first determines whether it works for you:
| Limit type | What it looks like | Who it hurts |
|---|---|---|
| Requests per minute / day | A hard 429 once you cross it, resets on a fixed window | Anything user-facing or in a loop |
| Tokens per minute / day | Long prompts burn the quota in a handful of calls | Anything with large context — RAG, code review, long chat |
| Model access | Only small or older models, or a single vendor | Anyone who needs a frontier model for reasoning quality |
| Data terms | Inputs may be used for training, or logged | Anything with customer data or proprietary code |
The first two are visible in a dashboard. The last two are the ones people discover too late.
These are the recurring names in the free-tier conversation, and what each is genuinely good for:
Fast inference on open-weight models with a published free rate limit. The draw is latency, not capability: tokens per second is unusually high, which makes it good for interactive demos and bad for anything that needs Claude- or GPT-level reasoning. Rate limits are enforced per minute and per day, so a burst burns through them quickly.
Similar shape: free tier on their own hardware, very high throughput, a daily token allowance that resets. Good for trying large-volume batch work at zero cost; not a substitute for a frontier model on hard tasks.
Free access to Gemini models through an API key, with rate limits and — importantly — terms that differ from the paid API. Worth reading before you send anything you would not want used as training data. Useful for prototyping Gemini-specific features without a billing account.
:free variantsRoutes to models with a :free suffix, subject to availability and its own rate caps. Convenient because it is one endpoint for many models; the catch is that free variants are the first thing to be throttled or withdrawn when capacity is tight, so they are unreliable as a production dependency.
People search for a "free API key" and land on pages that list a number without the two things that decide whether it is usable: whether a card is required, and what the free allowance actually throttles. Here is the same comparison with those two columns, because they are the ones that end the experiment.
| Provider | Free allowance | Card required | What really limits you |
|---|---|---|---|
| OpenAI API | No standing free tier for new API accounts | Yes, to enable paid usage | Billing setup is the gate, not the quota |
| Anthropic API | No standing free tier; credits appear through programs and grants | Yes for paid usage | Access to credits, then per-model rate limits |
| Google AI Studio (Gemini API) | Free tier with per-minute and per-day request limits | No | Rate limits, and free-tier data may be used to improve products |
| Groq | Free tier with request and token rate limits | No | Requests per minute on the models you actually want |
| Cerebras | Free tier with request rate limits | No | Rate limits; model list is narrower than the paid catalog |
| OpenRouter | A rotating set of :free model variants | No | Which models are free changes; free routes are rate-limited |
| Tapix AI | $1 trial credit on signup, spendable on the full catalog | No, not to start | Only the $1 itself - no rate limit below the per-key concurrency cap |
Read the card column first. A "free" key that requires a card is not free for someone without one - it is a trial with a payment method attached. That is the single most common reason a developer gives up on a provider before writing a line of code.
Two more things worth checking before you build on a free tier: whether the free tier is allowed for commercial use, and whether your prompts can be used to improve the provider's models. Both are usually answered in a sentence buried in the terms, and both have cost teams a rewrite.
max_tokens: 16. No key, no experiment.Free becomes expensive at the moment your time costs more than the balance would. Three concrete signals:
At that point a prepaid balance of a few dollars buys back the thing you were actually short of, which is iteration speed.
A reseller is not free and never will be. What it does is remove the rate ceiling: you pay per token, so the limit becomes budget rather than quota, and budget is something you control. What it does not do is make a frontier model cheap — nobody does that, and a provider claiming otherwise is either subsidising or substituting.
The honest comparison is therefore not "free vs paid" but "quota-limited free vs budget-limited paid". If your workload fits inside a free tier's limits, use the free tier. If it does not, the question becomes which paid option has predictable arithmetic — which is what the three checks are for.
We are the paid option in this comparison, and we are not the right answer for everyone. If a free tier covers your workload, use it — we say that in the pricing breakdown too. What we offer is a prepaid balance from $1 with no card, per-token billing with no rate ceiling of our own, and published prices you can check against the vendor's list on every row of the pricing page.
If you want to test the arithmetic before committing anything, the cost calculator takes your token counts and gives you the monthly number.
Compare prices to official list · Verify a provider in five minutes · All guides
Keep reading: Trial without a card · What $1 actually buys · Cheap LLM API options