Free OpenAI-compatible LLM APIs

Free tiers you can call with the OpenAI SDK by changing the base URL and the model name. Each base URL below was read in the provider’s documentation on the date shown.

Confirmed in the documentation (6)

Free LLM APIs with a confirmed OpenAI-compatible endpoint
Provider	OpenAI SDK	Free allowance	Requests / min	Requests / day
GroqFree tier	Yes No logprobs, logit_bias or n > 1. 3 Oct	30 req/min · 1,000 req/day Chat models; 8K tokens/min and 200K tokens/day. 3 Oct	30 openai/gpt-oss-120b, openai/gpt-oss-20b, qwen/qwen3.8-27b. 3 Oct	1,000 Chat models. Prompt-guard models: 14,400/day. 3 Oct
Google Gemini APIFree tier	Yes 3 Oct	Not published Limits are per project and shown only inside Google AI Studio. 3 Oct	Not published Not published per model: “Rate limits … can be viewed in Google AI Studio.” 3 Oct	Not published Not published per model: “Rate limits … can be viewed in Google AI Studio.” 3 Oct
OpenRouterFree tier	Yes 3 Oct	20 req/min · 50 req/day 1,000 req/day once the account has bought at least 10 credits. 3 Oct	20 All free model variants together. 3 Oct	50 1,000/day once the account has bought at least 10 credits. 3 Oct
Cloudflare Workers AIFree tier	Yes Chat completions and embeddings; the base URL contains your account id. 3 Oct	10,000 Neurons/day Resets at 00:00 UTC. Going above it needs Workers Paid ($0.011 per 1,000 Neurons). 3 Oct	Not published The allowance is counted in Neurons, not in requests. 3 Oct	Not published The allowance is counted in Neurons, not in requests. 3 Oct
Hugging Face Inference ProvidersFree tier	Yes The docs use the OpenAI client against the router URL. 3 Oct	$0.10 of credits/month Listed as “subject to change”. Past it, usage needs purchased credits. 3 Oct	Not published The allowance is a monthly credit, not a request count. 3 Oct	Not published The allowance is a monthly credit, not a request count. 3 Oct
CerebrasTrial only	Yes 3 Oct	$5 trial credit, 30 days Granted “after adding a verified payment method”. 3 Oct	5 Free Trial: gpt-oss-120b and qwen-3.8-27b. 3 Oct	Not published The Free Trial table lists no requests-per-day figure. 3 Oct

55 of 72 cells above were read on the provider’s own page in the last 30 days (oldest check: 3 Oct 2026). Each check mark opens that page. The other 17 say “not verified” instead of repeating a number from someone else’s article.

Not confirmed (1)

We did not read a compatibility page for these providers within the last 30 days.

NVIDIA NIM

Base URLs and a model to try

OpenAI-compatible base URLs per provider
Provider	base_url	Model id to try	Key variable	Read on
Groq	https://api.groq.com/openai/v1	openai/gpt-oss-120b	GROQ_API_KEY	3 Oct
Google Gemini API	https://generativelanguage.googleapis.com/v1beta/openai/	gemini-3.8-flash	GEMINI_API_KEY	3 Oct
OpenRouter	https://openrouter.ai/api/v1	openrouter/free	OPENROUTER_API_KEY	3 Oct
Cloudflare Workers AI	https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1	@cf/meta/llama-3.1-8b-instruct	CLOUDFLARE_API_TOKEN	3 Oct
Hugging Face Inference Providers	https://router.huggingface.co/v1	deepseek-ai/DeepSeek-V3-0324	HF_TOKEN	3 Oct
Cerebras	https://api.cerebras.ai/v1	qwen-3.8-27b	CEREBRAS_API_KEY	3 Oct

The same request on each of them

base_url: https://api.groq.com/openai/v1
model: openai/gpt-oss-120b
key from: $GROQ_API_KEY

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.groq.com/openai/v1",
    api_key=os.environ["GROQ_API_KEY"],
)

reply = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[{"role": "user", "content": "Reply with the word: ok"}],
)
print(reply.choices[0].message.content)

Endpoint read on 3 Oct 2026· console.groq.com

The key is read from an environment variable on your machine. This site has no field for keys and never asks for one.

What to check before you rely on it

Compatible is not identical

Every provider implements the chat completions route; fewer implement everything around it. Groq documents that logprobs, logit_bias and n above 1 are not supported. Cloudflare Workers AI offers chat completions and embeddings on its compatible endpoint. If your app uses tool calling, JSON mode or streaming, test those on the model you picked, because support is per model.

Put three values in configuration

The base URL, the model id and the key are what change between providers. With those three outside the code, moving from one free tier to another, or to a paid tier when the free one ends, is a configuration change.

Cloudflare’s URL contains your account

The Workers AI base URL includes the account id, so it is different for every user. Replace {account_id} with the id shown in your dashboard, and authenticate with an API token.

Model ids are not portable

The same open model has a different id on each provider: openai/gpt-oss-120b on Groq and gpt-oss-120b on Cerebras. Keep the id next to the base URL.

Questions and answers

What does “OpenAI-compatible” mean?

The provider exposes the same HTTP routes and JSON shapes as the OpenAI API, mainly /chat/completions. You keep the OpenAI SDK and change two things: the base URL and the model name.

Which free LLM APIs are OpenAI-compatible?

Confirmed in their documentation on 3 Oct 2026: Groq, Google Gemini API, OpenRouter, Cloudflare Workers AI, Hugging Face Inference Providers. Cerebras is also compatible, but its free offer is a 30-day trial.

Is compatibility complete?

No. Groq documents that logprobs, logit_bias and n greater than 1 are not supported. Cloudflare Workers AI supports chat completions and embeddings on its compatible endpoint. Test the parameters your app depends on.

Can I switch providers without changing code?

Mostly: put the base URL, the model name and the key in configuration. Rate-limit errors and model ids differ, so keep the retry logic and the model name configurable too.