Most “free LLM API” tiers are not unlimited access; they are a rate-limited model, a daily request quota, or a referral credit. This comparison covers six providers, each with a verified free-tier mechanism and a specific number attached, so you can see exactly what “free” buys before you build against it.
Quick Comparison of Free LLM APIs
| Provider | Free-tier type | Concrete limit | Model coverage |
|---|---|---|---|
| Novita AI | Referral credit (both sides) | $10 in credit to each of the referrer and the new referred account per successful referral, up to $500 total | Full serverless LLM API catalog |
| OpenRouter | Rate-limited free model variants | 20 requests/minute; 50 requests/day without purchased credits, 1,000 requests/day with $10+ lifetime credits | 17 of 414 catalog models carry a :free variant |
| Google Gemini API | Free-of-charge flagship tokens; grounding requires billing | $0 per-token input/output/context caching on Gemini 3.7 Flash with no billing enabled; Grounding with Google Search is not available until billing is enabled, then capped at 5,000 requests/month | Gemini 3.x model family |
| Groq | Rate-limited free tier, per model | 30 requests/minute, 1,000 requests/day, 200K tokens/day on openai/gpt-oss-120b | 13 hosted models across production and preview tiers |
| Cloudflare Workers AI | Daily compute allowance | 10,000 Neurons/day at no charge on both Free and Paid Workers plans | 84 hosted models as of 2026-08-18 (catalog updates continuously) |
| Hugging Face Inference Providers | Monthly credit | $0.10/month for Free-tier accounts | Hundreds of models across 18 partner inference providers |
Key Takeaways
- No provider on this list gives unlimited free LLM API access. Every free tier is bounded by a request rate, a daily cap, or a dollar credit.
- Credit-based free tiers (Novita AI, Hugging Face) do not cap requests per minute; rate-limited free tiers (OpenRouter, Groq) do, and the cap is model-specific.
- OpenRouter’s
:freevariants exist for a minority of its catalog (17 of 414 models at the time this was checked), not the whole marketplace. - Google’s free tier is model- and feature-specific: token pricing on the current flagship Flash model is $0 with no billing enabled, but Search grounding is not part of that no-billing tier at all — it only becomes available once you enable billing, at its own separate monthly cap.
- Free-tier numbers change without much notice. Every figure below has a source and a check date; re-verify before you build a production dependency on any of them.
How We Compared These Free LLM APIs
Most existing “free LLM API” roundups either skip concrete numbers, skip Novita AI entirely, or go stale between updates. This comparison uses three checks for every row:
- A stated free-tier mechanism. Each provider’s free access falls into one of three forms: a one-time signup or referral credit, a rate-limited free model, or a daily compute/request allowance.
- A verified numeric limit. Every request-per-minute, request-per-day, dollar amount, or token cap is pulled from the provider’s own current pricing or rate-limit documentation, not a secondhand blog post.
- Live status at check time. Free-tier terms shift often (OpenRouter’s threshold for higher daily limits, Google’s per-model free pricing, and Cloudflare’s Neuron allocation have all changed in past cycles), so every number below is dated.
Providers are listed by the rough size of their daily free allowance where one is published, with credit-based tiers (which are not capped by request count) placed by their dollar ceiling instead.
Free LLM APIs Compared in Detail
Cloudflare Workers AI: largest raw daily allowance
Cloudflare’s Workers AI gives every account, including the Free Workers plan, a flat daily allocation of Neurons rather than a per-model request cap. At the time this was checked, that allocation was 10,000 Neurons per day at no charge, with usage above that billed at $0.011 per 1,000 Neurons on the Paid plan. Neuron cost varies by model, so 10,000 Neurons translates to a different number of actual requests depending on which model you call.
Best for: teams already running on Cloudflare Workers who want inference bundled into the same platform. Watch for: the free allowance is Neurons, not requests, so a large or long-context model burns through it faster than a small one.
OpenRouter: broadest catalog, narrow free slice
OpenRouter routes to hundreds of models from many providers, but only a subset carry a :free suffix. At the time this was checked, 17 of 414 catalog models had a free variant. Free variants are capped at 20 requests per minute; the daily cap depends on your account’s lifetime credit purchases, 50 requests per day if you have purchased less than $10 in credits total, rising to 1,000 requests per day once you have purchased at least $10.
Best for: testing a specific open-weight model that happens to have a :free variant, without committing to a provider.
Watch for: the free slice is a small fraction of the catalog, and the useful daily cap requires a one-time $10 credit purchase, not just a signup.
Groq: fastest inference, smallest catalog
Groq’s free tier is rate-limited per model rather than account-wide. For openai/gpt-oss-120b, the published limits are 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute, and 200,000 tokens per day. Other hosted models on Groq’s free tier carry their own separate limits.
Best for: low-latency prototyping where Groq’s inference speed matters more than model breadth. Watch for: limits are set per model, so switching models mid-project can mean re-reading the limits table.
Google Gemini API: free flagship tokens, narrow grounding cap
Gemini’s free tier is priced per model and per feature rather than as one flat allowance. On the current flagship Gemini 3.7 Flash model, input price, output price, and context-caching price are all listed as free of charge as long as billing is not enabled on the account. Grounding with Google Search, a separate add-on feature, is not available at all with no billing enabled; it only becomes usable once you turn on billing, at which point the paid tier includes 5,000 free grounding requests per month shared across the whole Gemini 3.x model family, then billed at $14 per 1,000 requests.
Best for: teams that want $0 token cost on a current flagship model and do not need Search grounding without enabling billing. Watch for: the token-level free pricing and Search grounding are not the same tier — grounding requires billing to be turned on before any request allowance exists, so you cannot get grounded answers on a pure no-billing account.
Hugging Face Inference Providers: smallest credit, lowest friction
Every Hugging Face account gets a monthly credit toward Inference Providers, a marketplace that routes to third-party inference backends for open models. Free-tier accounts get $0.10 per month; PRO accounts get $2.00 per month, and Team or Enterprise accounts get $2.00 per seat.
Best for: quick one-off calls against an open model already hosted through Hugging Face’s provider marketplace. Watch for: $0.10 per month covers a small number of calls on most models; it resets monthly but does not roll over or stack.
Novita AI: no rate-limited free model, but a real referral credit
Novita AI’s serverless LLM API catalog is priced per model, and no current model in that catalog is offered at $0 per token. The free-tier mechanism instead comes from Novita’s referral program: a successful referral gives both the referrer and the new referred account $10 in LLM API credits, up to $500 in total credits for a single referrer. That credit applies against Novita’s full serverless model catalog rather than a single free-tier model, so it is not capped by requests per minute or per day the way OpenRouter’s or Groq’s free tiers are.
Best for: developers who want free usage against production-grade model pricing and availability, rather than a separate, more limited free-tier model. Watch for: the credit is referral-gated, not an automatic grant on every new signup, and it is a spending allowance rather than an unlimited free model; once it is spent, standard per-token pricing applies. Always check Novita’s current pricing page for live rates before estimating how far a given credit will go.
How to Choose the Right Free LLM API
Match the free-tier mechanism to what you’re actually doing:
- Running a lot of small test calls against one open model: a rate-limited free tier (OpenRouter’s
:freevariants or Groq) gets you real request volume without spending anything, as long as the specific model you need has a free variant. - Prototyping against a production model you might ship with: a referral credit (Novita AI) lets you spend against the same catalog and pricing you would use in production, instead of switching models when you graduate off a free tier.
- Testing multiple providers or models quickly: Hugging Face’s Inference Providers credit is small but requires the least setup if you already have an account.
- You need search-grounded answers, not just raw model calls: check Google’s grounding cap specifically. It is not included in the no-billing free tier at all — you must enable billing before any grounding requests count against a (paid-tier) monthly cap.
Whichever tier you pick, read the provider’s own current rate-limit or pricing page before you build; every number above has a check date, and free-tier terms are among the most frequently revised pages on any inference platform.
Who This Comparison Is Not For
None of the tiers above are built to carry production traffic. If your app already needs more than a few thousand requests a day on a consistent basis, a support SLA, or access to closed frontier models like GPT-5 or Claude, none of the six providers here offer that on a free plan — you are looking at paid usage regardless of which one you pick. This comparison is also not a substitute for reading a provider’s own current terms before you ship: it tells you what each free tier looked like on the check date, not what it will look like when you actually need the capacity.
Using Free Credits on Novita AI
To use Novita AI’s referral credit:
- Create a Novita AI account via a friend’s referral link, or invite a friend using your own link, so both sides receive the $10 referral credit.
- Generate an API key from the Novita AI console.
- Call any model in the serverless LLM API catalog through the OpenAI-compatible endpoint at
https://api.novita.ai/openai/v1/chat/completions, using your API key and a model ID from the current catalog. - Share your referral link from the Novita AI referral page to add $10 in credit per successful referral, up to $500 total.
curl https://api.novita.ai/openai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <NOVITA_API_KEY>" \
-d '{
"model": "<MODEL_ID>",
"messages": [{"role": "user", "content": "Hello"}],
"max_tokens": 512
}'Check the Novita AI model catalog for current model IDs and per-model pricing before spending credit against a specific model.
Pricing, License, and Availability Notes
Free-tier terms in this comparison were checked on 2026-08-18 directly against each provider’s own pricing or rate-limit documentation. Signup credits, referral caps, rate limits, and daily allowances are the kind of detail providers change with normal product updates, not always with a changelog entry. Re-check the source page for any provider before relying on a specific number in production planning, and treat every figure above as a snapshot rather than a permanent guarantee.
Decision Checklist for Developers
- Confirm which free-tier form applies (signup credit, referral credit, rate-limited free model, or daily compute allowance).
- Confirm the model you actually need is covered by that free tier, not just the provider in general.
- Confirm the specific numeric limit (requests/minute, requests/day, tokens/day, or dollar credit) on the provider’s current documentation, not this article, before shipping anything that depends on it.
- Confirm whether the free tier requires a purchased credit threshold before higher limits apply (as OpenRouter’s does).
- Plan for what happens when the free tier runs out: does pricing simply kick in, or does the request get rejected?
FAQ
Is there a truly unlimited free LLM API?
No. Every provider compared here bounds free usage by a rate limit, a daily cap, or a dollar credit. Providers marketed as “free” are describing a capped tier, not unlimited access.
Which free LLM API has the highest request-per-day limit?
Among the rate-limited free tiers checked here, OpenRouter and Groq both publish a 1,000 requests/day ceiling on their respective free-tier models, once OpenRouter’s $10 lifetime credit threshold is met. Cloudflare Workers AI’s 10,000 Neurons/day allowance is a compute budget rather than a fixed request count, so it does not translate to a single comparable request number.
Does Novita AI offer a free LLM model?
Novita AI’s current serverless LLM API catalog does not list a model priced at $0 per token. Its free-tier mechanism is a referral program that gives $10 in credit to both the referrer and the new referred account per successful referral, up to $500 total, spendable against the full serverless catalog.
Do free-tier limits apply per API key or per account?
This varies by provider and is documented on each provider’s own rate-limit page; Groq, for example, states that its limits apply at the organization level rather than per individual API key. Check the specific provider’s documentation for the account-versus-key distinction before assuming request pooling behavior.
