What a micro-SaaS actually costs to run
A paying AI app on the default indie stack — Vercel, Supabase, Resend, Stripe, plus an LLM API — costs about $25.88/month at 100 users, $78.78 at 1,000 and $653 at 10,000, assuming 5% of users pay $29/month. Every figure below is computed from the providers' published tiers (re-verified 2026-10-24 or later) and the index's cross-checked LLM prices — and you can re-run it on your own numbers.
The bill at 100, 1,000 and 10,000 users
| Line | 100 users | 1,000 users | 10,000 users |
|---|---|---|---|
| LLM API (DeepSeek V4.1 Flash at Baidu) | $0.17 | $1.73 | $17.29 |
| Hosting — Vercel Pro | $20.00 | $20.00 | $20.00 |
| Database — Supabase | $0.00 | $0.00 | $25.00 |
| Email — Resend | $0.00 | $0.00 | $20.00 |
| Payments — Stripe (2.9% + $0.30) | $5.71 | $57.05 | $571 |
| Total / month | $25.88 | $78.78 | $653 |
Assumptions: 25 AI requests/user/month (chat profile: 1,200 in / 400 out tokens, 60% cached), 12 pageviews/user, 3 transactional emails/user, 500 KB stored/user, 5% paying $29/mo. Medians, not measurements — adjust every one of them in the calculator.
The LLM line, at three levels
The model is the one line you choose, and it moves the bill more than anything else. Three levels, taken from the Epoch Capabilities Index rather than picked by hand, each at its cheapest host, on the same chat profile:
| Level | Model | Epoch score | 100 users | 1,000 users | 10,000 users |
|---|---|---|---|---|---|
| Best score per dollar | Mistral Nemo at Io Net | 119 | $0.06 | $0.61 | $6.10 |
| Cheapest in the top half of scores · in the bill above | DeepSeek V4.1 Flash at Baidu | 156 | $0.17 | $1.73 | $17.29 |
| Cheapest in the top tenth | GPT-6 Sol at OpenAI | 164 | $12.76 | $128 | $1,276 |
Where the money actually goes
Three things decide the bill, and none of them is what most builders expect. First, the floor is $20/month, not $0: Vercel's Hobby plan is non-commercial only, so the moment you charge users you are on Pro — the free tiers of Supabase (0.5 GB, 50k auth users) and Resend (3,000 emails/month) hold much longer. Second, the LLM line is yours to size: a mid-range model stays a small line at 1,000 users, a top-tier one can become the biggest, which is why the LLM calculator ranks all tracked offers on your own workload. Third, past roughly 1,000 users your biggest line is Stripe, because payment fees scale with revenue rather than traffic — at 10,000 users the 2.9% + $0.30 per transaction comes to $571/month. That one you should be glad to pay.
When each free tier runs out
Under the same assumptions: Supabase's free tier is exhausted by database storage around 1,000 users (0.5 GB at 500 KB/user) long before its 50k-user auth limit; Resend's 3,000-email tier lasts to exactly 1,000 users at 3 emails/user; and Vercel Hobby was never available to a commercial app in the first place. The calculator computes these thresholds live for whichever stack you pick — including Netlify, Cloudflare, Railway, Neon, Turso, PlanetScale, Postmark, SES and merchant-of-record payment options.
→ Run your own numbers in the stack cost calculator
FAQ
How much does it cost to run an AI micro-SaaS with 1,000 users?
Around $78.78 per month for a typical stack — Vercel Pro ($20.00), Supabase ($0.00), Resend ($0.00), the LLM API ($1.73 on DeepSeek V4.1 Flash at Baidu, a mid-range model) and Stripe fees ($57.05 on 5% of users paying $29/mo). Payment processing, which scales with revenue, is usually the biggest line — a good problem to have.
What is the minimum monthly cost of a commercial app on this stack?
About $20/month. Vercel's free Hobby plan is restricted to non-commercial use, so any app that charges users starts on Pro at $20/seat. Database, email and the LLM API can all sit in free tiers at low traffic; payment fees only exist once you have revenue.
Is the LLM API the expensive part of running an AI app?
It depends on the model more than on anything else. At 1,000 users making 25 AI requests each per month, the LLM line is $0.61 on Mistral Nemo at Io Net (best score per dollar), $1.73 on DeepSeek V4.1 Flash at Baidu (cheapest in the top half of scores), $128 on GPT-6 Sol at OpenAI (cheapest in the top tenth). The cheap end stays under the hosting bill; the top end can become the biggest line.
Sources: each provider's published pricing page, re-verified by hand on 2026-10-24 or later (linked per line in the calculator and in the raw dataset /api/stack-pricing.json, CC BY 4.0). LLM prices from the index's daily cross-checked pipeline.