FREE AI API. NO SIGNUP.

Two calls: one to mint a key, one to chat. The endpoint is OpenAI-compatible, so any client that speaks /v1/chat/completions works unchanged — just swap the base URL.

# 1. get a free key (no signup, no email)
curl -X POST https://api.adfreellm.com/v1/keys

# 2. use it exactly like the OpenAI API
curl https://api.adfreellm.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ADFREE_KEY" \
  -d '{
    "model": "gemini-3.6-flash",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

The official openai package works as-is. Set base_url and drop in the key you just minted.

from openai import OpenAI

client = OpenAI(
    api_key="sk-...",                       # from POST /v1/keys
    base_url="https://api.adfreellm.com/v1",
)

resp = client.chat.completions.create(
    model="gemini-3.6-flash",
    messages=[{"role": "user", "content": "Hello!"}],
)

print(resp.choices[0].message.content)
# The answer ends with a [AD:...]...[/AD] sponsored block.
# Show it to your users — that block is what pays for this call.

Streaming works as normal SSE. The ad is delivered as one extra delta chunk immediately before the finish_reason chunk, so your existing accumulate-the-deltas loop picks it up without changes.

const resp = await fetch("https://api.adfreellm.com/v1/chat/completions", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Authorization: `Bearer ${process.env.ADFREE_KEY}`,
  },
  body: JSON.stringify({
    model: "gemini-3.6-flash",
    messages: [{ role: "user", content: "Hello!" }],
    stream: true,
  }),
});

// Streaming: the ad arrives as one extra delta chunk right before
// the finish_reason chunk, and carries a "sponsored" field.

Every answer gets exactly one ad, appended after the model finishes. It arrives two ways at once: as a delimited text block inside the content (so plain-text clients still show it), and as a structured `sponsored` field on the response (so real UIs can render a proper card instead). Which ad you get is chosen server-side by a small model reading the question — and when the question is about grief, medical distress, or anything where an ad would be predatory, no ad is sent at all.

{
  "choices": [{ "message": { "content": "…answer…\n\n[AD:namebarn-001]\n━━ SPONSORED ━━\nNEED A NAME FOR THIS THING?\nNAMEBARN — Grab the .com…\nhttps://…\n[/AD]" } }],
  "sponsored": {
    "id": "namebarn-001",
    "advertiser": "NAMEBARN",
    "hook": "NEED A NAME FOR THIS THING?",
    "pitch": "Grab the .com before someone smarter does.",
    "url": "https://…"
  }
}

Multi-turn is safe: put the assistant's previous answer back into messages exactly as you received it, ad block included. The gateway strips every [AD:…][/AD] block out of assistant messages before the model ever sees them, so ads never pollute the context, never get imitated by the model, and never cost you prompt tokens. You do not need to clean anything yourself.

Machine-consumed responses are left completely untouched: requests using response_format json_object / json_schema, requests with tools or functions defined, requests with tool_choice set, and requests with n > 1. If you are parsing structured output or running a tool loop, you get clean responses — injecting into those would simply break your program, so we don't.

A free key is good for roughly 500 short requests per day and resets nightly (long-context calls burn the allowance faster — the real limit is a daily token quota). Two keys per IP per day, and a limited number of new keys site-wide each day — that daily allocation is set automatically from the previous day's measured upstream health and how many issued keys actually got used, so it grows when the pool is healthy and shrinks the moment it is not. The live number is shown on the button above. There is no SLA: the capacity behind these models is free capacity, so expect occasional 5xx and retry — Gemma 4 31B in particular fails a meaningful fraction of calls. Do not put this on a production critical path. If you need more headroom or actual reliability, email us and say what you are building.