Disclosure: this post links to AICredits with an affiliate link. If you sign up through it, MasterPrompting may earn a commission. The free options below cost nothing and I name them first.
You can call a hosted LLM for free today from at least five providers, but each free tier has a different catch: a daily request cap, a monthly call cap, a data-use clause, or a model list that changes under you. This post lists what each vendor's own page said when I read it on 2026-10-23, so you can pick one in ten minutes instead of finding the catch in production.
I did not call any of these APIs in this session. The limits and terms below come from documentation pages only. The code is syntax-checked but unrun.
What do the free LLM API tiers say today?
Where a page showed no free-plan number, the table says so rather than guessing.
| Provider | What the page states (read 2026-10-23) | Prompt use stated |
|---|---|---|
| Google Gemini API | Free tier exists. Limits are measured in requests per minute, input tokens per minute and requests per day, per project, with daily quotas resetting at midnight Pacific. The rate-limit page gave no free-tier numbers; check AI Studio. | Unpaid services: content may be used to improve products, and human reviewers may read it. Paid: not used. |
| Groq | The limits page I could read showed only Developer plan figures (for example openai/gpt-oss-120b: 30 RPM, 1K RPD, 8K TPM, 200K TPD). Free plan numbers are in your account's limits page. | Not checked. |
| OpenRouter | Models with an ID ending in :free: 20 requests per minute; 50 requests per day if you have bought fewer than 10 credits ever, 1,000 per day if you have bought at least 10. | Page mentions separate settings for paid and free models; it does not say what free-model providers do with prompts. |
| Cloudflare Workers AI | 10,000 Neurons per day at no charge on both Workers Free and Paid plans, reset at 00:00 UTC. | Not checked. |
| Cohere | Trial keys: 1,000 API calls a month; chat models capped at 20 requests per minute. | Not checked. |
| Mistral | Docs say free mode gives "included monthly usage within the limits shown on the Limits page" of the admin panel. No numbers on the docs page. | Docs link to terms and privacy policy only; not read. |
| GitHub Models | Retired. See below. | Not applicable. |
Two things stand out. First, only Gemini's page told me plainly that free usage can be used to improve the vendor's products. For the others I could not confirm either way, which is not the same as "safe". Second, the numbers are small. A 50-requests-a-day cap is a toy; a 1,000-call monthly cap is a demo.
GitHub Models is gone
If an older tutorial tells you to use a GitHub token against GitHub Models, check first. GitHub's docs page says that as of July 30, 2026 "the playground, model catalog, inference API, and bring your own key (BYOK) are no longer available to any customer" and points to Azure AI Foundry and GitHub Copilot. I could not find a free API path on that page.
Pages I could not use
Mistral's help-centre and tier pages returned 404 or no figures, so I cannot give you its free-tier numbers. Check admin.mistral.ai after signing up. Groq's free-plan limits are only visible inside its console.
Which models can you actually call for free?
Model names rotate quickly, so treat this as a snapshot. Google's pricing page on the same date showed "free of charge" input and output for, among others, Gemini 3.8 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, Gemini 2.5 Pro and Gemma 4. It showed Gemini 3.1 Pro Preview and Gemini 4 Argon as not available on the free tier. If your code pins a model name, expect to edit it every few months. The earlier Gemini Flash free tier guide covers set-up in India in more depth, though its model names and prices predate this check.
How do you call all of them with one Python snippet?
Gemini, Groq and Cloudflare each document an OpenAI-compatible endpoint, so one client works for all three. The base URLs and model strings below are copied from their docs pages of 2026-10-23 (Gemini's own example uses gemini-3.8-flash; the Groq model ID comes from its limits table; Cloudflare's endpoint notes list @cf/openai/gpt-oss-120b).
import os
from openai import OpenAI
PROVIDERS = {
"gemini": {
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
"key_env": "GEMINI_API_KEY",
"model": "gemini-3.8-flash",
},
"groq": {
"base_url": "https://api.groq.com/openai/v1",
"key_env": "GROQ_API_KEY",
"model": "openai/gpt-oss-120b",
},
"cloudflare": {
"base_url": "https://api.cloudflare.com/client/v4/accounts/"
+ os.environ.get("CLOUDFLARE_ACCOUNT_ID", "ACCOUNT_ID")
+ "/ai/v1",
"key_env": "CLOUDFLARE_API_TOKEN",
"model": "@cf/openai/gpt-oss-120b",
},
}
def ask(provider: str, prompt: str) -> str:
cfg = PROVIDERS[provider]
client = OpenAI(base_url=cfg["base_url"], api_key=os.environ[cfg["key_env"]])
resp = client.chat.completions.create(
model=cfg["model"],
messages=[{"role": "user", "content": prompt}],
max_tokens=200,
)
return resp.choices[0].message.content
if __name__ == "__main__":
print(ask("groq", "Reply with the single word: ready"))
Status: syntax-checked with py_compile, not run, because I had no keys in this session. Whether a given free key can use that model is account-dependent.
OpenRouter works the same way with https://openrouter.ai/api/v1 as the base URL and a model ID ending in :free; I did not re-read that base URL today, so confirm it in their quickstart. Cohere is not in the snippet because I did not verify an OpenAI-compatible endpoint for it.
How do you stay under a free tier's limits?
Every page above counts requests, tokens or both, and any one limit can trip you. Gemini's page says that exceeding any single dimension triggers an error even if you are under the others, and that limits are per project, not per key. So creating more keys in one project does nothing.
Practical habits:
- Retry with exponential backoff on 429s. The rate limits and backoff guide has a tested pattern.
- Cache identical prompts locally while developing.
- Keep a daily counter. OpenRouter's 50-per-day cap can disappear in one test loop.
- Log which provider answered, so a quality change after a model rotation is traceable.
When should you move off the free tier?
Move when one of these is true:
- A limit blocks real users. Daily caps reset on the vendor's clock, not yours.
- Your data matters. Gemini's terms put the line at paid versus unpaid: paid usage is not used to improve products. For other vendors, read the terms before sending anything real.
- You want one key for many vendors. Juggling six free keys is its own maintenance cost.
For the data point, you can simply enable billing with the vendor. If you are in India and want rupee billing, there are gateways. OpenRouter is the well-known non-affiliate one, but its FAQ lists cards, AliPay and crypto and not UPI, per our AICredits vs OpenRouter comparison (read 2026-10-16).
AICredits is the one I would look at for UPI: per its docs you top up in INR via UPI, the minimum is Rs 50, and the endpoint is OpenAI-compatible at https://api.aicredits.in/v1, so the snippet above only needs a new entry in PROVIDERS. Its docs describe a 5% forex buffer and a 5% platform fee, plus about 2.36% on the top-up itself. Read those lines on their site and run your own numbers; the AICredits review walks through them. Pay-by-card direct to the vendors, or a card with low forex charges, are the alternatives. Our UPI payment guide lists the options.
A five-minute decision rule
- Learning or a weekend demo, no private data: Gemini or Groq free tier.
- Need an open model on your own Cloudflare account: Workers AI's 10,000 daily Neurons.
- Just trying many models: OpenRouter
:freeIDs, accepting the daily cap. - Anything with real user data: pay, even Rs 50 at a time, and read the data terms first.
I would skip building a product on any free tier. They are for finding out whether the idea works.
