DEV Community

Cover image for $0/Month, 5 AI Apps, All on Gemini: Here's Exactly What the Free Tier Gives You (and What Breaks First)
S M Tahosin
S M Tahosin

Posted on

$0/Month, 5 AI Apps, All on Gemini: Here's Exactly What the Free Tier Gives You (and What Breaks First)

#webdev #ai #googleaichallenge #opensource

Every "build AI apps for free" tutorial stops at localhost. Nobody writes the next post — the one where you actually ship the thing, users show up, and you find out what the free tier really means.

I spent the last four months shipping five AI products on Google's Gemini free tier. All of them are live right now, all of them still cost me $0/month, and I can tell you exactly where the bottom starts to fall out.

This is the post I wish I'd had in January.

The 5 apps (all live, all free to use, all open-source)

App What it does Live URL Stack
HOCKS AI ChatGPT clone + image analysis + video analysis + website generator hocks.app React + Vite + Firebase + Gemini + OpenRouter
MaxAI Writer 6 free AI writing tools (resume, cover letter, email, blog) maxai-writer.pages.dev Next.js + Cloudflare Pages + Gemini
PromptCraft AI Prompt generator for Midjourney / DALL-E / Stable Diffusion promptcraft-ai.pages.dev Next.js + Cloudflare Pages + Gemini
EcoSense AI Upload a photo, get a carbon footprint estimate in 60 seconds ecosense-ai.pages.dev Next.js + Cloudflare Pages + Gemini Vision
EcoBot OpenClaw skill for personal carbon tracking & green-living tips GitHub OpenClaw + Gemini

Every link works. Click any of them — they'll load in under 2 seconds and respond to your first prompt in about the same.

The architecture (one diagram you can steal today) 

                     ┌─────────────────────────────────┐
                     │           Your User              │
                     └─────────────┬───────────────────┘
                                   │
                     ┌─────────────▼───────────────────┐
                     │    Cloudflare Pages (free)       │
                     │    Unlimited bandwidth           │
                     │    500 deploys/mo                │
                     │    Next.js SSR at the edge       │
                     └─────────────┬───────────────────┘
                                   │
                     ┌─────────────▼───────────────────┐
                     │   @google/generative-ai (SDK)    │
                     │   Runs in edge runtime           │
                     └─────────────┬───────────────────┘
                                   │
     ┌─────────────────────────────┼─────────────────────────────┐
     │                             │                             │
┌────▼─────────┐         ┌────────▼─────────┐         ┌─────────▼────────┐
│ Gemini 2.5   │         │  Gemini Vision   │         │  Imagen 3        │
│ Flash        │         │  (multi-modal)   │         │  (image gen)     │
│ 60 rpm free  │         │  same quota      │         │  throttled       │
└──────────────┘         └──────────────────┘         └──────────────────┘
Enter fullscreen mode Exit fullscreen mode

That's the whole thing for 4 out of 5 apps. No backend server, no database (for the text tools), no paid auth, no paid CDN. The API key sits in an environment variable on Cloudflare. The app calls Gemini. Gemini returns text or an image. Done.

HOCKS AI is the exception — it has chat history, image uploads, and a user system, so it uses Firebase (also free: Spark plan gives 50K monthly active users, 1 GB storage, 10 GB bandwidth).

The exact cost breakdown

Here's what I actually burn against each free tier in a normal month:

Service Free ceiling What I use Headroom
Cloudflare Pages Unlimited bandwidth, 500 builds/mo ~60 builds ✅ comfortable
Cloudflare Workers (edge runtime) 100k requests/day ~3–5k/day peak ✅ huge margin
Firebase Spark (HOCKS AI only) 1 GB storage, 10 GB egress, 50k MAU < 200 MB, < 2 GB, ~400 MAU ✅ not even close
Gemini 2.5 Flash (text) 1,500 requests/day, 1M tokens/min ~600 req/day peak ✅ enough
Gemini Vision Shared with Flash quota low volume ✅ fine
Imagen 3 Per-region throttled ~30 gens/day ⚠️ tight
Total monthly bill $0.00

My only cash outlay is a $5/month GCP e2-micro VM that runs a completely separate project (an n8n-based social-media autopilot). That's unrelated to the five apps above — it's just the only recurring AI-adjacent bill I have.

What actually breaks first — the stuff nobody tells you

This is the part the free-tier content farms leave out.

1. The free Gemini quota changes without warning

In January, gemini-1.5-flash gave you 1,500 requests/day free. By February some regions got bumped down to 100/day on certain models while gemini-2.5-flash took over the bigger quota. I had live apps calling the wrong model and hitting 429s within an hour of the change.

Fix: Never pin to a specific model in your .env. Use an env var + a server-side fallback chain:

// lib/gemini.ts
const MODELS = [
  process.env.GEMINI_MODEL ?? "gemini-2.5-flash",
  "gemini-1.5-flash",
  "gemini-1.5-flash-8b",
];

async function callGemini(prompt: string) {
  for (const model of MODELS) {
    try {
      return await genAI.getGenerativeModel({ model }).generateContent(prompt);
    } catch (e: any) {
      if (e.status !== 429 && e.status !== 503) throw e;
      // fall through to next model
    }
  }
  throw new Error("All Gemini models rate-limited");
}
Enter fullscreen mode Exit fullscreen mode

This single pattern has saved me 3 outages so far.

2. Cold starts on serverless functions kill perceived free

When Cloudflare Workers is your only compute, cold starts are fine (sub-100 ms). When you move a single endpoint to Firebase Functions (e.g., for server-side auth verification), you'll see 1–3 second cold starts on low-traffic endpoints. Your free tier is fine. Your users perceive the product as slow.

Fix: Keep everything at the edge when humanly possible. Firebase Functions only for the 2–3 endpoints that truly need Node runtime (e.g., Firebase Admin SDK calls that can't run at the edge). And keep the edge function warm with a 5-minute cron that pings it (cron job is free too).

3. Imagen hits the wall way before you think

Text generation on Gemini Flash is genuinely abundant. Imagen is not. One creative user running an image-generation tool can burn through your daily quota in a single sitting. I've had PromptCraft AI go 429 at 11 am on a Tuesday because someone discovered it on Reddit.

Fix: Per-user rate limiting via a KV store (Cloudflare KV is free up to 100k reads/day), with a friendly "come back in an hour" message. And NEVER enable anonymous image generation without it.

// middleware.ts
const key = `ratelimit:${clientIp}:${today}`;
const count = (await env.KV.get(key)) ?? 0;
if (Number(count) >= 10) {
  return Response.json({ error: "daily_limit", retry_after: "tomorrow" }, { status: 429 });
}
await env.KV.put(key, String(Number(count) + 1), { expirationTtl: 86400 });
Enter fullscreen mode Exit fullscreen mode

The uncomfortable honesty

"Free" is not free. You pay in:

  • Architecture constraints. You can't build a Stable-Diffusion-grade image tool on Gemini's Imagen free tier. You can't run a vector DB with millions of embeddings on the free Pinecone plan. Your design choices shrink.
  • Vendor risk. Google can and will change the quota. I've rewritten prompts three times already because a model got retired mid-quarter.
  • Perceived quality. Users don't care that your stack is free. If a cold start ruins their first click, they're gone.

So "free" really means: I pay in time instead of dollars, and I architect around the limits instead of buying my way past them.

That's a fair trade when you're shipping side projects solo. It's a terrible trade if you have a paying customer who needs 99.9% uptime.

The code

All five are open-source. Steal whatever's useful:

The Next.js-on-Cloudflare-Pages-plus-Gemini template is effectively the same package.json across 4 of the 5 repos. If you clone any one and swap the prompts + UI, you have your own free AI app in an afternoon.

What I'd do differently

  1. Build the rate-limit and quota-fallback before the first public URL, not after the first 429. I learned this the painful way on three of the five apps.
  2. Pick one analytics vendor early (Cloudflare's built-in analytics is free and enough). Mixing Google Analytics + Vercel Analytics + a custom event table was a mistake I still haven't finished untangling.
  3. Write one shared @you/ai-utils package for prompt templates and Gemini fallbacks. I copy-pasted the same fallback chain five times before I learned my lesson.

What's next

I'm about to add a 6th: a fully open-source social-media autopilot that posts to six platforms on a 12-hour cron, also on free tiers. Repo is at https://github.com/x-tahosin/social-media-autopilot if you want to watch that one go up.

If any of this was useful, a heart + a follow pushes me up in the Google AI Challenge rankings, which genuinely would make my week. And if you ship something based on any of these five — drop me the link, I read every one.

Built with Gemini. Hosted for $0. Written because nobody else was writing it.

Top comments (0)