The Tool

I built a Reddit Keyword Research Tool — it analyzes Reddit posts and comments to surface high-value keywords for content creators, marketers, and SEO professionals.

The stack:

• Frontend + Hosting — Vercel

• AI Engine — Anthropic Claude API

• Caching — Upstash Redis

• Built with — Claude Code

Live, functional, and serving real users. Here's how much it costs to run.

The Infrastructure Cost Breakdown

Yes, a live AI-powered SaaS tool for under $2/month. Here's exactly how.

Vercel — Free Tier is More Than Enough

Vercel's Hobby plan gives you:

• 100 GB Fast Data Transfer

• 10 GB Fast Origin Transfer

• 1M Edge Requests/month

My current usage after launch:

Fast Data Transfer:  15.19 MB / 100 GB   ✅
Fast Origin Transfer: 1.51 MB / 10 GB    ✅
Edge Requests:        1.4K / 1M          ✅

For a tool in early stages, you won't come close to these limits.

Note: Vercel's Hobby plan prohibits commercial use. Once you start monetizing, upgrade to Pro ($20/month). For a solo project, that's still very manageable.

Upstash Redis — The Free Tier Workhorse

Upstash gives you:

• 10,000 commands/day

• 256 MB storage

• 1 free database

I use Redis for caching Reddit API responses. This alone significantly reduces how many times the Claude API gets called — saving money on both ends.

Critical tip: Always set a TTL (time-to-live) on your cached keys:

# Cache Reddit results for 1 hour
redis.setex("keyword:python", 3600, json.dumps(results))

Without TTL, your 256 MB fills up and caching breaks silently.

The Biggest Mistake — Wrong Claude Model

This one cost me money unnecessarily.

My tool was running on Claude Opus — Anthropic's most powerful and most expensive model. For a keyword research tool that analyzes text patterns, Opus is massive overkill.

Model comparison:

A single line change fixed it:

# Before — expensive
model="claude-opus-4-..."

# After — 20x cheaper, same quality for this use case
model="claude-haiku-4-5-20251001"

Rule of thumb: Match the model to the task complexity. Keyword extraction and text analysis doesn't need Opus-level intelligence.

Protecting Yourself — Spend Limits

Anthropic's API default spend limit is $500/month — which sounds fine until you realize a misconfigured loop or unexpected traffic spike could drain your credits fast.

Two things I set up immediately:

1. Reduced monthly spend limit from \(500 → \)5

   • Go to: platform.anthropic.com → Settings → Limits → Adjust limit

2. Added email notification at $3 threshold

   • You get alerted before you run out, not after

This is especially important because the API uses prepaid credits — once your credit runs out, the tool stops working. Better to know at $3 than to find out when users start complaining.

What I Did NOT Do (But Should)

Prompt Caching — Anthropic offers prompt caching that can reduce input token costs by 50–90% for repeated system prompts. My tool uses the same system prompt on every call. Enabling this is next on my list.

Summary — The Optimization Checklist

✅ Use Vercel Hobby tier for early-stage tools
✅ Cache API responses with Upstash Redis (always set TTL)
✅ Pick the right Claude model — Haiku for most tasks, Sonnet for complex ones
✅ Set a spend limit well below your credit balance
✅ Add an email alert at 80% of your credit
✅ Keep auto-reload OFF until you're generating revenue

The Numbers

Starting cost: ~\(1.64/month (on the wrong model)
After optimization: ~\)0.10–0.20/month (estimated at Haiku rates)
Remaining credit: $3.38
Projected runway: 15–30 months

What's Next

Once the tool gains traction:

• Enable prompt caching

• Move to Vercel Pro for commercial use

• Add a paid tier — even $5/month from 10 users covers 2+ years of API costs

Building lean from day one is not just about saving money. It's about buying yourself time to find product-market fit without burning budget on infrastructure.

Built using Claude Code. Hosted on Vercel. Cached with Upstash. Running on almost nothing.

Follow along as I build more tools and document what actually works.

Try the tool here: [https://keywords.innerfuel.live/\]